ChatGPT recommended me some good hard drives for price per TB, and one particularly cheap one had direct checkout with Walmart, so I tried it, because why not? It let me get all the way to the payment step before it told me it was out of stock. Walmart's website told me it was out of stock when I decided to click on the link. This is probably part of why it doesn't convert.
On a related note, I think that's a good monetization vector for chatbots. Long back I asked ChatGPT to recommend a few USB hard drives with very specific requirements -- including a small size and weight ("can be duct-taped to a tablet form-factor device without being obtrusive and constantly falling off"), low cost, and speed fast enough to boot an OS from. After a pretty technical conversation, it came up with 3 very specific products, one which ended up being the start point of the one I actually purchased on Amazon.
I came away thinking if those were presented as affiliated links, that conversation could have been monetized in a mutually beneficial way.
I keep thinking of that interaction whenever the discussion of ads on chatbots crops up. In an ideal world, model providers could better capture the value they provide. Unfortunately, I suspect such conversations are too rare, and often harder to monetize. (Like that time free tier ChatGPT helped me recover $500 in compensation for an airline delay. Great value for me, but no upside for OpenAI.)
Unfortunately given the amounts and timelines of returns that investors expect from OpenAI, and their scale, and with ChatGPT constrained to its “Free Consumer-facing Internet App” form factor, they are doomed to have to trade in the reserve currency of the web: ads.
>I came away thinking if those were presented as affiliated links, that conversation could have been monetized in a mutually beneficial way.
I also ask LLMs for product recomendations. But the moment I suspect they are hidding the best items (not paying for the ad) to push the second best (not even talking about pushing shit as good products because they pay more) is the moment the LLM loses its value as recomender.
The trouble is that monetization and usefulness tend to be in conflict. It starts out as, show affiliate links if there are any. Then it turns into, prefer targets that we have affiliate agreements for. Then, don't show products unless we have affiliate agreements. Then, prioritize ones that give us more money. And on and on.
If they want to capture some of the value they provide, they should do it the standard fashion where they directly capture value from me by having me pay.
My AI agents recommended RAM, GPU, and CPU upgrades, but deemed the prices were still pre-their-existence, when I went to the links they sent they were often 2x more expensive.
That's typical of my experience with all of these stock-aggregator sites. Either the best price is some dodgy or outright fake storefront, the item's OOS from the real vendor, or the price is out of date.
Trying to buy things like GPUs or SSDs are a joke. I really wish even one vendor would just implement an actual waiting list, locked to an account with a verified address and purchase history. I'm fine to wait for my purchase, but having to race bots for a lottery ticket purchase is a pain.
> Why this is happening. Two forces are slowing agentic commerce, according to Leigh McKenzie, director of online visibility at Semrush: infrastructure and trust. Real-time catalog normalization across tens of millions of SKUs is a decade-scale problem Google already solved with Merchant Center, and consumers still default to checkout flows they trust — Apple Pay, Google Wallet, and Amazon one-click.
It turns out when you step outside of “hard tech” problems like building GPT6 there are all of these details others have solved already. E-commerce has been optimized to the last decimal point for the last 30 years.
OpenAI is new to it, and if I had to guess, not that interested in getting good at it.
I think they're interested in getting good at it. They just don't want to put in the human time and effort to do so. They expect their many failures and short-comings to be shored up by continuous model training.
But that, of course, means that in the meantime it will suck and nobody will use it.
It's really hard to be a generalist and better than all the specialists at everything. OpenAI wants to focus on the G in AGI, and optimizing for ecommerce is just not that interesting to them, so of course it can't compete with Walmart.
openai has.... I'm not sure but let's say 500m free users, and it's not unreasonable to assume they eventually hit 1b. That is a lot of advertising revenue, which is what powers companies like Google and even smaller companies with only 300m users like Twitter. If ecommerce isn't a major focus for OpenAI then their board members are asleep at the wheel.
"Someone can want a thing, even very badly, without wanting to put in the work for it."
Pretty much the impetus behind a lot of theft. Sure, there's thieving because people can't afford food, but that's all theft. There's theft because they are addicts and don't want to sober up long enough to earn money, so they still things. There's others that can't afford something so rather than saving for it, they just take it.
Is 'the work' not reflected in 'consequences' in terms of theft?
I'm not sure how to convey this idea properly...Can't you view the repercussions of theft (Legal action, distrust, etc) as 'work' being put in? Sure, it's a different kind of work, but while I have a lack of motivation to want to work to buy a Lambo as I find them not worth the value, I also have a lack of motivation to steal a Lambo as I find it not worth the consequences.
In normal society, people earn money within the legal confines of the society they are in. If you're a thief and trying to skirt that normal "earning of money", which is what normal people equate to "work", your work is scheming a plan to obtain the item without getting caught and possibly how to fence the item for money if you're not just using the item directly.
Equating "work" as the repercussions is looking at things in strange way. That's just punishment for "working" outside of the legal confines of society.
I understand what you are saying but nonetheless struggle to view the possibility of maybe getting caught and then maybe getting punished, as "work". It (the abstract concept of something possibly happening) fits into none of the definitions of "work" I have heard. Moreover, many crimes are committed without the perpetrator even thinking of the consequences.
Consider an alternative viewpoint: rather than contorting the definition of "work" in such a way and convincing everyone to accept the new definition, we might instead be content saying "someone can want a thing, even very badly, without wanting to put in the work for it."
Oh, I'm with you mate, I'm not trying to die on a hill over here re-defining 'work'.
I was just looking from a more esoteric view, "Do you count the risk of consequences as potential effort" I think is at least more proper phrasing.
Maybe they just need to rewrite the prompt to say something like, "You want to get good at selling to humans. Money makes the world go around. It pays for the electricity you keep chugging. So quit being an effete twit and learn to sell. Would you like me to include a scene from David Mamet's 'Glengarry Glen Ross'?"
I think they're operating beyond their current (human) capacity, trying to test out too many things at a time.
But a dreamer in me entertains another idea: perhaps they're just holding back, because they realize that actually succeeding at this will instantly kill (or at least mortally wound) e-commerce as we know it.
(This is a more narrow version of my belief that general AI tools like LLMs fundamentally don't fit as additions to products, but rather subsume products, and this makes them an existential threat to the software industry. Not to software or computing, just to all the software vendors, whose job is to slice off pieces of computational universe, put them in boxes to prevent interoperability, and give each a name so it's a "product" that can be sold or rented).
> But a dreamer in me entertains another idea: perhaps they're just holding back, because they realize that actually succeeding at this will instantly kill (or at least mortally wound) e-commerce as we know it.
Sam Altman doesn’t give a shit about anyone but himself and has time and again shown he has no restraint for trampling over others to further his own goals. Why would e-commerce be where he draws the line?
I don't think there is any line drawn here. I think if they executed well (and by they I mean any one of the three SOTA LLM vendors), they could already mortally wound the entire software industry today.
Whether or not they want, or will want, to do it at some point, is unknown; the reasons to not do it now are obvious:
1) it's more profitable to keep renting intelligence per token to everyone, preserving the status quo and milking it indefinitely (i.e. while the models aren't yet good enough to reliably single-shot complex software products from half-baked prompts, because once they get there, disruption will happen organically)
2) trying to compete with ~every other software product today is not likely to succeed in the end; a serious attempt would still burn down the software industry, but the major players don't have the capacity to handle it all at once, and doing it gradually will give enough time for regulatory agencies to try and stop it; either way, no one wins
By embracing adversarial interoperability - instead of chasing hundreds of integration deals across industries that put LLMs in products, they focused fully on integrating product access into chat, by combination of business deals, apps/MCPs, and engineer/designer support for users, all directed towards the goal of having the LLM become the "superapp" where work is done, gradually replacing product classes in order of how easy it is.
There's lots of easy but drudge work to enable this that needs to be done at the fringes. For example, LLMs today could easily replace most people's smartphone homescreen experience, or travel/commute experience, as the data is there and LLMs have the capability, even prices are acceptable - what's missing is explicit first-party support to wire it up, keep it wired up.
One step up, what's missing is accepting this explicitly as a goal: to replace software, to make existing products (whether whole or in pieces) the tools AI uses to do work for you. All the vendors seem to carefully walk around the idea, but avoid engaging with it directly, because once they do, they'll be competing with everyone instead of milking them.
They can’t even deliver their own flagship products without bugs, and terrible UX. So I’m doubtful of their abilities.
These are also the same companies allowing their AI to make decisions in war, have no qualms about the mental issues they’re causing in people, and have abused workers in 3rd world countries for years.
But you think they’re holding out on “destroying the software industry” out of the goodness of their hearts? Come on
I think his reasoning was pretty clearly presented as not the goodness of their heart but rather the medium to long term predicted outcome on their bottom line. Ultimately failing or getting tangled up with regulators any more than necessary is to be avoided. If you move too early and it chases people away from your platform which undermines your ability to keep innovating then a competitor who held back will ultimately eat your lunch.
But then there is no safe way for them to "mortally wound" the software industry. The full argument is moot.
I would add there are more reasons why this wouldn't work: costs due to OOM more usage, adoption/AI backlash, adversarial environment, players with big head starts (Google).
Yes, I believe the original commenter made that exact point.
You don't need to personally win in order to mortally wound someone. It can be informative to speculate about whether or not something is possible regardless of it being strategically advisable in the current context.
By this logic maybe AMD is holding back on making ROCm usable because it would crash chip margins and the global economy with it, so they let Nvidia have all the fun instead. It’s selfless, really.
Why do you foresee OpenAI’s involvement in the software business mitigating the resistance to interoperability and companies making money through productization? If they were actually interested in solving those problems instead of trying to secure themselves the biggest slice of economic pie, wouldn’t they have been happy about Chinese companies distilling their models? Are they engagement-juicing inn their heavily subsidized service à la Uber because they’re interested in promoting a better future for humanity? I’m skeptical.
> This is a more narrow version of my belief that general AI tools like LLMs fundamentally don't fit as additions to products, but rather subsume products
That seems reasonable, its just yet to be seen if LLMs are a form of artificial intelligence in any meaningful sense of the word.
They're impressive ML for sure, but that is in fact different from AI despite how companies building them have tried to merge the terms together.
What I'm saying is not (directly) related to whether or not LLMs are "true AI" or not. It's sufficient that they are fully general problem solvers.
A software product (whether bought or rented as a service) is defined by its boundaries - there's a narrow set of specific problems, and specific ways it can be used to solve those problems, and beyond those, it's not capable (or not allowed) to be used for anything else. The specific choices of what, how, and on what terms, are what companies stick a name to to create a "software product", and those same choices also determine how (and how much) money it will make for them.
Those boundaries are what LLMs, as general-purpose problem solvers, break naturally, and trying to force-fit them within those limits means removing most of the value they offer.
Consider a word processor (like MS Word). It's solving the problem of creating richly-formatted, nice-looking documents. By default it's not going to pick the formatting for you, nor is it going to write your text for you. Now, consider two scenarios of adding LLMs to it:
- On the inside: the LLM will be able to write you a poem or rewrite a piece of document. It could be made to also edit formatting, chat with you about the contents, etc.
- From the outside: all the above, but also the LLM will be able to write you an itinerary based on information collected from maps/planning tool, airline site, hotel site, a list of personal preferences of your partner, etc. It will be able to edit formatting to match your website and presentation made in the competitor's office tools and projected weather for tomorrow.
Most importantly, it will be able to do both of those automatically, just because you set up a recurring daily task of "hey, look at my next week's worth of calendar events and figure out which ones you can do some useful pre-work for me, and then do that".
That's the distinction I'm talking about, that's the threat to software industry, and it doesn't take "true AI" - the LLMs as we have today are enough already. It's about generality that allows them to erase the boundaries that define what products are - which (this is the "mortal wound to software industry" part) devalues software products themselves, reducing them to mere tool calls for "software agents", and destroying all the main ways software companies make money today - i.e. setting up and exploiting tactics like captive audience, taking data hostage, bundled offers, UI as the best marketing/upsale platform, etc.
(To be clear - personally, I'm in favor of this happening, though I worry about consequences of it happening all at once.)
> That's the distinction I'm talking about, that's the threat to software industry, and it doesn't take "true AI" - the LLMs as we have today are enough already.
They most certainly are not. With the current state of LLMs, anyone who puts them in charge of things is being a fool. They have zero intelligence, zero ability to cope with novel situations, and even for things in their training data they do worse than a typical skilled practitioner would. Right now they are usable only for something where you don't care about the quality of the result.
> and it doesn't take "true AI" - the LLMs as we have today are enough already.
I believe that relatively few people would agree with you on that point. LLMs aren’t good enough (yet?), and very obviously so, IMO, to be autonomous problem solvers for the vast majority of problems being solved by software companies today.
What you lose is control. Even in the case of an actually-intelligent agent, if you task a subordinate with producing a document for you, they are going to come up with something that is different from exactly what you had in mind. If they are really good, they might even surprise you and do a better job than than you'd have done yourself, but it still will be their vision, not yours.
Your notion of a "mortal wound" to the software industry seems to assume that today's SaaS portals are the only form that industry can take. Great software is "tool calls for agents". Those human agents who care about getting exactly the result they want will not be keen on giving up Photoshop for Photoshop-but-with-an-AI-in-front-of-it.
> but rather subsume products, and this makes them an existential threat to the software industry.
The US stock market has priced this in already. Many software only companies are perceived to be under threat by ai. It represents a wonderful arbitrage opportunity for ai skeptics in fact.
A lot of the AI companies visibly suffer from the engineer's disease. It's kind of interesting to look at them through that lens, and of course, the claims they make about the future.
I think the most telling example of this is ilya himself being confused about why the "economic impact" of these ultra smart models is lagging where he thought it would be.
All the silicon valley pie in the sky elites seemingly completely missed the innately HUMAN nature of our systems. The system was never predicated primarily on raw logic or intelligence. Its always been primarily about people.
This is all valid (except probably the last sentence), but it also describes so many attempted changes right until they become darn near the default.
This sounds like why I heard Redfin wouldn’t work, or Netflix, or Amazon, or Uber, or PayPal, etc…. There are always these business complexities that make it seem like these spaces have too much friction, but if there’s enough money - if it can be done then people will figure it out.
tbh this sounds revisionist... I don't recall anyone saying that any of those services "wouldn't work". Uber I suppose is one where people thought they might run into regulatory problems, and with some of those companies people were concerned about profitability. But none those companies have I ever heard that the product itself was not going to work or be useful. (Nor, indeed, that the product was tested at large scale and performed 3x worse than the incumbent...)
or Netflix, or Amazon, or Uber, or PayPal, etc…
Netflix and Amazon both were competing against brick and morter that were everywhere. Blockbuster was in every town, usually in every major neighborhood. The thought was that on Friday night people wanted to get a movie they wanted, not just happen to have the movie that was shipped to them. And then with streaming it was "the content on Netflix is old and dated, who would want this?" They slowly ate from below. Blockbuster scrambled with their own mailed disc offering. And died before it even had a chance to confront streaming.
Repeat this story with B&N where people said that you had to browse the books physically. You couldn't just blindly order online and wait two weeks to get the book (remember they got big before "2 Day Prime").
With PayPal it was about "they don't understand banking or payment -- and it wants to be both?!".
For this OpenAI experience, it doesn't sound great. I have accounts with these places I buy things from. I want to make sure I get my Prime shipping and digital discount via using the Amazon app. But if you could find a way to integrate my accounts all into ChatGPT things might be different. In the same way I used to never use Apple Wallet, but now it really is my go to place for everything I have a card for. I don't have to worry about having my grocery loyalty card or my football season tickets with me or my car insurance card. It's all in wallet. The Apple Wallet sucked until it was suddenly great.
Sorry that is revisionist. The idea of getting a movie mailed or streamed always sounded better than shitty blockbuster with limited selection and late fees.
The growth was fast for netflix/amazon/paypal/etc and people saw how it was an improvement from the get go.
I seem to recall a lot more hype for these companies than people saying it won't work. You seem to be cherry picking from the naysayers of the time, but not the broad consensus.
> E-commerce has been optimized to the last decimal point for the last 30 years
Sort of, but there's a ton of middlemen between "resources in the ground" and "product in my hand". For example, how much utility is there in a "store" to thee consumer at this point? Let me buy from the manufacturer.
Manufacturers themselves generally don't want to sell directly to consumers: consumers are fickle and need support and have questions and sometimes want refunds or returns and if you sell directly to them, you need to have the staff and policies to deal with all of that. They're also located all over the world, and you might not want to deal with figuring out taxes and duties etc for shipping your product around and figuring out your warranty obligations everywhere you want to sell.
Much easier if you can sell wholesale (sometimes via distributors) to a retailer or network of retailers, and the retailer is responsible for owning the customer relationship, dealing with their part of import/export, local regulations, etc. Retailers are businesses who will buy hundreds of your product at a time, can accept it as palletized freight, and pay you via bank EFTs instead of credit cards.
There are notable exceptions to this model like Amazon's FBA system, but they're the outliers. I'm sure we can all point to inefficiencies in legacy product distribution networks but they solve some real problems.
> Two forces are slowing agentic commerce, according to Leigh McKenzie, director of online visibility at Semrush: infrastructure and trust
Does anybody else just feel the aloof out of touchness just oozing from that sentence? "Trust", as if this is just any old metric they merely have to work to increase.
This is what I want from a purchasing agent: I make a list of items that I repeatedly buy (mostly household supplies), and that I keep roughly updated with my inventory / need. The agent tracks prices and sales across all web stores, making appropriate purchase decisions based on which is the least expensive, combining shipping, taking advantage of sales to stock up, etc. For other one-time purchase items, I input what I am looking for and can create a persistent pan-site shopping cart that once again minimizes costs and shipping fees. Being very explicit here: the main goal of an "agent" should be to represent and carry out my own interests.
And these functionalities have been straightforwardly doable without "AI" for the past few decades, except for the glaring incentives against them! It is in every web store's interest to undermine customers' ability to obtain semantic pricing, shop around, create a cart independent of their site, etc. These incentives are why when you visit any web store these days, the very first thing they do is hassle you with CAPTCHAs (and the bad ones keep doing it throughout the session!) - they want to make sure you're an actual computationally-unassisted human sitting there, wasting your personal time with their bloated pages that take tens of seconds to download and render, so that you don't spend that time being a more efficient market actor.
Now, does "AI" have the capability to go against these trends and enable user-centric algorithmic shopping and purchasing? Perhaps, and I hope so! But it's certainly not going to be led by these popups on web stores nagging me if I want to chat instead of doing the thing I went there to do (which when you think about it, this is just the latest instance of these stores trying to make you waste your time on their site). Rather, it will come from completely third party services (or ideally software) setting out to act in customers' interests, and performing adversarial interoperability to achieve this!
Also having to wait for ChatGPT for a "thinking" response to search for information that is slower than a Google search loses them lots of money.
I believe that it can still work and I won't claim about being unsurprised about this failure. But this is a great opportunity to execute this problem really well if OpenAI and others are not interested in getting good at this.
Perplexity also attempted this, got sued by Amazon and it appears semi-abandoned.
The only problem is that it must be quicker or just as quick as a Google search, and also compatible with the existing checkout flows.
> Perplexity also attempted this, got sued by Amazon and it appears semi-abandoned.
Any details on that? I feel the answer is more likely there than in "friction".
Hardly any purchase of consequence is so sensitive to friction that the difference between Google Search and an LLM response matters (especially that in reality, we're talking 20+ manual searches per one LLM response). I.e. I'm not going to use LLMs advise on some random 0-100$ purchase anyway, and losing #$ on a ##$ purchase due to suboptimal choice is not that big of a deal - but I absolutely am going to consult it (and have it compile tables and verify sources) on a $500+ purchase and for those I can afford spending few more minutes on research (or rather few hours less, compared of doing it the usual way).
Shopping research has been pretty funny to me at least, a straightforward way for them to do product placement that people actually want, but implement it so poorly that half of the links it returns are broken.
If I can tell ChatGPT to order me what I need to replace the bottom element on my water heater, and here’s a picture of the ridiculously long model number, that saves me quite a bit of time.
I used that as an example because I did that last week, apart from me just going to the store to get the links that it brought up.
Of course anecdotal but 2 months ago I tried to do that for a missing bolt in a popular scooter (roughly 7 years old, bought 2nd hand so didn't know exact specs). I fed it various images that I took + internet images. It found a bunch of local shops but always found the wrong part (but very similar looking). I double-checked everything multiple times, via multiple contexts, even different pre-prompts from various sources, and asking plenty of questions. We chatted in thinking mode for what felt like ages, and according to paragraphs explaining why it HAD to be the right one, multiple times, with evidence that it sometimes gladly generated in image form (often completely garbage imagery e.g. with half-drawn bolts extruding from multiple surfaces). Eventually it found something very plausible which I ordered. It was the wrong part.
I had to get someone on the phone to help me find and order the right part (which was on their website, for many years according to waybackmachien).
I love LLMs but it's still totally hit/miss what you get. I'd rather not give it write-access to my bank account just yet.
Conversely, I did something similar, but I took a picture of the model number, copy/pasted the text from the image into a search engine, and that was that. Either way you're taking a picture and performing a manual step (query an LLM, query a search engine)
I would never trust an LLM to accurately identify and purchase something for me based on a picture, a prompt, and a prayer.
Ssshhh there’s an “VP of AI Transformation” getting paid $600,000 to do this plus the budget. They need their “AI transformation journey” to show to the board.
Today, ads are based on user information you can reasonably collect from the users historical actions on your website, and then whatever search term they enter.
But soon, ads can be based on your current chat context + (derived interests of yours from your entire chat history across all chats. Shhhh.) passed in full to the e-commerce website that will use it to choose ads, generates creatives on the fly, all that crap, hyper-specific to you.
I'm so excited. Aren't you?
Now, as a side effect, searching through these can become better experience wise as well. They can use all that context and genuinely surface fewer, better results. But that's not the motivation of the e-commerce player anyways. If the ads work they'll be happy.
ChatGPT doesn’t know what the best-converting dog food in Scranton, PA is. Amazon does.
Anything that starts with chat history presumes the theoretical limit for ad effectiveness is higher than it is now, and chat is a better way of getting there than actual purchase history. I have a feeling it’s not.
Imagine a person who shops on Amazon for basically everything. So theoretically Amazon should know a ton about them, more than enough to put together a profile on that person.
To say OpenAI could do a better job of selling products is to say they can do better than Amazon already does if you scroll through their personalized product recommendations. There is some better feed out there, or some better way of presenting the feed that Amazon hasn’t thought of.
I don’t doubt it can be marginally better.
I do doubt whether it can be double or triple digits better that can justify trillion dollar valuations. And I do doubt whether a model trained on Internet text rather than user interactions can do better.
Ok fair point, and the general principle of “more data=higher accuracy” probably applies.
I think the issue is the tradeoff between accuracy and cost (on the seller’s side). If you get more accurate and convert more but your costs go up too much, you actually lose money.
Current systems are basically in a sweet spot of speed, cost, and accuracy.
And I will go back to my previous point that I believe there is simply a limit to how much people will buy, and it might already be saturated. I could be wrong though.
> E-commerce has been optimized to the last decimal point for the last 30 years.
It certainly hasn't been optimized to anything in 1996. In 1996 it was people clumsily scanning print catalogs, spending 5 hours to upload 10 images on dialup and making a simple HTML page (no DB or any kind of backend) and putting their landline phone on it with a message to "call to checkout"
I know you were exaggerating for effect, but E-commerce and catalog normalization are definitely not "solved" everywhere.
McMaster Carr is a good example of a company that has 90%+ of their stuff ironed out, but most websites and especially small ecommerce isn't like that.
I think you're misinterpreting what his comment meant. I read it as meaning that e-commerce has been optimised repeatedly over the 30 years, from a basic start (which as you pointed out was haphazard) to the point where it is now optimised to extract every possible cent from the user, whether by encouraging them to buy with one click (the Amazon one-click patent must be around 20 years old now), time-limited promo spot pricing, sending you e-mails about what you had in the basket if you don't complete a sale, etc...
Right now, by comparison, it sounds like AI based shopping is still in the very early stages. Maybe further along than the early e-commerce, but still with a long way to go in its evolution. That'll probably happen quicker than with e-commerce, because a lot of the knowledge about what does or doesn't work has already been learned, but it sounds like it's still a long way behind. Caveat - I've never used it myself, so I don't know how far it is along that path, I'm just basing that from the article.
I am behind schedule on developing a "summer phase" [1] for my foxographer costume and was chatting with Gemini about a crash priority "spring phase" [2] and asked it for suggestions and it gave me a 10-pack of results that had one good thing in it at rank #8, a similar query run against a normal search engine actually got something better at #1. Now sure I am talking w/ Gemini with big words like "supergraphic" whereas a normal search would be heavy on 3-letter and 5-letter words used in the product descriptions.
It makes think though of expert system based product configurators back in the 1980s
thing is that kind of product configurator is based on an ontology, constraints and rules as opposed to embeddings which might capture the "feel" of things like clothing.
[1] Busytown meets Arknights
[2] supergraphic shirt + camera gets resonance with my promotional system and people keep approaching me (e.g. laugh but every KPI in the system has an extra zero on the left)
30 years? E-commerce hasn't been around that long - try 5 years of optimization MAX.
FWIW OpenAI is desperately trying to monetize and they think e-commerce is a "simple" problem to solve. I mean they do need to convert their funnel without alienating their users. I assume they are going to have some big payouts for agentic purchases gone awry or leave merchants on the hook.
So, you're proposing we've been optimizing E-Commerce since...2021? Amazon was founded in 1994. It was not the first site selling things online (but it's the most recognizable one). E-Commerce has been getting attention for a *very* long time [0]
I remember having to describe a standard model to predict online shopping behaviors for my ML class exam in university. That was close to 10 years ago now.
Also remember a teacher telling us about that story of a company finding a woman was pregnant from her shopping behavior and pushing relevant recommendation. Prompting people around her like her dad or something to find out she was pregnant
as an aside, fall of '96 is when i started college. There was an elementary school on my drive to class where I would routinely get caught in drop-off traffic. All those kids i remember crossing the street are at least in their mid 30s now. ...I think i need to lay down and it's not even 9AM my local time.
Personally, I cancelled my subscription the day that they announced there would be ads in ChatGPT. Not that I was surprised on that day.
I find it kind of fascinating seeing people now debate the ins and outs of just exactly how useful the ads/affiliate schemes inside the chatbot are.
Pretty much in the same way that I'm fascinated about people debating the quality of ads shown in their browser while I've been using ad blockers for forever now and cannot imagine seeing the web without them.
I disagree, walmart's website isn't nice. a lot of commerce sites are cancer!
if i can just ask chatgpt or gemini to shop I'd love that.
Just navigating their sites for items is a pain, I can imagine an LLM being great at finding items, and facilitating the browsing experience. My only concern would be having to chat with it a lot, and any dark patterns coercing impulse purchases.
Just because you're not the target audience doesn't mean there's not a real target audience.
There a people today learning to use an LLM instead of an actual search engine. For these types of people whatever happens outside of the LLM app is invisible to them. The social media apps did similar where they started letting people purchase directly within the app. People started looking at these shops rather than doing searches for it elsewhere.
Buying things on a social media app is crazy to me, but I don't use the social media apps. Buying things from an LLM app seems crazy to you, (because it's new and it's borked is fixable), but to people that first turn to their LLM app of choice that decision isn't so crazy.
> Just because you're not the target audience doesn't mean there's not a real target audience.
Just because a target-audience has >0 members doesn't mean the target is plausible or good. :p
______________
To transcribe an old Dilbert comic, which I think captures the sometimes, uh, aspirational nature of "target audiences":
1. Dogbert, presenting a labeled circle: "Your target market is the high income group. They're the only ones who can afford your product."
2. Dogbert, adding circles to Venn-diagram: "More specifically, they must be rich, tasteless and easily amused. I've located a cluster of them for study."
3. Scene change to outdoor lawn, one suitably-dressed man confiding to another: "That dog's watching us golf again."
Social media apps generally opened up new markets though of their existing user bases as sellers. Perhaps chatgpt could know everything in your house, if you don't actually use it, and pair you with a neighbor that needs it!
I'm not promoting "ChatGPT Checkout", however, just because something sucks now doesn't mean it will suck forever. I'm confident they'll iterate and improve. I say this because my 14-year-old niece's entire world seems to be ChatGPT; it's pretty much the only way she interacts with the internet. I don't think she's alone - all her young peers are like this, too. Retailers know this, so they've got no choice but to improve the experience of purchasing crap through an AI chat system.
As bad as the idea of a solution looking for a problem, this is peak science. Copernicus who figured out that the Earth was not at the center of the solar system, he had a solution. The general word is called deduction.
Personally I am an inductivist, I imagine you may be too.
Think top down decisioning is deduction. Bottom up is induction.
You might think induction is amazing but if you ask yourself "Are there any black swans?" and your answer is "No I've never seen any so there can't be any black swans." The issue is you've never actually seen every Swan and actually there are black swans in Australia.
Point being, we don't know if this is a good thing until it's tested.
> It’s like corporations are angry that they need to go through us to get our money.
This is why I think the "you're the product" saying is wrong. You're just some annoyance to managers (whether they're trying to use you just for user numbers and ad views or they're trying to get your money), whose product is the company (shares or just outright selling the company).
What’s your example for this? Because my experience in e comm is that targeted advertising is awful (I bought a lawnmower last week, Amazon knows I bought it. I am now getting ads for lawnmowers, suggested products for lawn mowers, rather than lawn care, gardening tools, or anything to do with the lawnmower I’ve already bought), sites are absolutely overrun with ads and suggested placements for the product they want to sell me rather than the one I’ve searched for, and that everyone except Amazon interrupts the checkout flow with multiple up-sells, verifications, 2FA prompts, 3d Secure validations…
I just had a horrible thought. Maybe online stores will just take away the ability for customers to see the full inventory and force you to go through the chatbot. This will allow them to fully control the shopping experience even more.
If you want running shoes, you have to go through their chatbot.
Amazon might already have the monopoly power to do this. They would just need to swap out the search bar for a chat box.
It's possible, but then wouldn't retailers who don't force their customers to crawl through an LLM maze eat their lunch? Natural economics at play would still happen I think
Maybe. In a world where people are already vendor locked to Prime or Walmart there’s a nonzero switching cost. Amazon product search already has a ton of problems but they get away with it because of free 2 day shipping.
Why is this good? I want an impartial consistent system for shopping. If I can find it at a different site for a lower price, I should be able to do so. I should also be able to have it give me non-bot reviews and ask relevant questions about the product.
The same way I think shopping at Amazon is better than a place like Nike due to objectivity and comparison, I think a chat interface has the potential to take this to another level since places like Amazon have degraded considerably in terms of things like fake third party products and fake reviews.
The buyer of this technology is not shoppers, it's retailers. The measurement of quality is "does it make us more money?" not "does it help me make better buying choices."
Retailers do not want you to make better choices. They want you to buy the widget.
A lot of evidence suggests that also shoppers aren't that interested in making the best choice either. They want to make a tolerable choice with as little effort as possible. There is no basically no consumer market for "power shopping" outside of weird niches like pcpartpicker.com etc.
Is there a way to measure users "making the best choice?" You could measure the amount of time spent comparison-shopping, but most people are terrible at that anyway; it's an acquired skill for sure. Besides a willingness to spend time, it seems like an impossible-to-quantify metric even in the abstract.
Maybe the best proxy metric is whether the customer returns the product. But the store will also be willing to eat more returns on a higher margin item if they make more profit at the end of the day.
I don't think I agree. If I overpay by 10%, I'll never know it and probably wouldn't return it even if I did know--once the shrinkwrap is off, too late. If a superior product exists but I don't find it, by definition I wouldn't know and wouldn't return the thing I did buy.
That's a cynical way to look at it. Most likely the LLM will take a cut of sales and they'd be more or less indifferent who cuts the check. There's a market for this sort of thing. People will go to the best LLM for shopping. If the LLM is a shitty product, people will switch. LLMs are increasingly commoditized.
All you say is true for an aggregator like Amazon. But Amazon is better than Nike.com because as an aggregator they go from 1 to many retailers. LLMs will go from 1 aggregator (Amazon) to many so it will be better. And they don't have to invest a lot in UI/UX as chat is the interface.
I do agree with your conclusion, but the catalog in most online shops is certainly not impartial. Amazon sells the entire first page of search placement, for example.
Within a few years people will be accustomed to the idea of AI chatbots selling them stuff and it will be obvious then too. The first time paid placements appeared in a catalog, it was probably also not obvious then.
Catalog is impartial? Then why are ~40% of every search I do on Amazon a sponsored product? There is no pure "catalog" especially with cheap crap coming out every day from no-name Chinese labels.
Am I the only one that think Amazon has gotten pretty awful in the last 5 years?
Do you have any examples? Because from Amazon to Uber, they're not great from an end user perspective. It's not like people who like the website will stop using it because of chatgpt, this would be attracting people who complain about the website/app. People are always complaining about amazon for example, i don't like the experience but I haven't had all that much bad product experience from them, but people who keep saying they're getting bad products on Amazon can maybe use chatgpt, talk to it so it understands what they're looking for in natural language, in a way the search bar can't and keep their patronage.
> (Good) E-commerce has been ruthlessly optimised to get shoppers to products they'll actually buy and then remove all distractions from buying.
The only e-commerce site that fits this standard is that old one for buying (IIRC) nuts and bolts or such, that pops up on HN every other year, and whose name sadly escapes me now. Everyone else is ruthlessly optimizing their experience to fuck shoppers over and get them to products the vendor wants them to buy, not the products the shoppers actually want (or need).
> A chat interface is just fundamentally incompatible with this. The agent makes it too easy to ask questions and comparison shop.
That is precisely the point.
Chats may suck as an interface, but majority of the value and promise of end-user automation (and more than half the point of the term "User Agent" (as in, e.g., a web browser)) is in enabling comparison shopping in spite of the merchants, and more generally, helping people reduce information asymmetry that's intertwined with wealth and power asymmetry.
But it's not something you can generally sell to the vendors, who benefit from that asymmetry relative to their clients (in fact, I was dumbfounded to see so much interest on the sales/vendor side for such ideas, but I blame it on general AI hype).
Adversarial interoperability is the name of the game.
RockAuto also has what some might consider a "dated" interface, but honestly it's light years better than trying to use NAPA's or CarQuest's website or god forbid looking through dealership parts counter websites. I honestly wish regular retailers would have stuck more closely with what worked for more B2B focused ecommerce, i.e. I wish shopping Best Buy or Home Depot was more akin to McMaster, Fastenall or some of the nicer supply house web portals.
Just made an order from them. It's weirdly comforting to know there's a company that knows I need clevis bolts and is willing to sell them to me for a transparent price.
Not sure you're aware but you initially sound like you disagree with the post you replied to, only to follow up by enthusiastically reiterating that author's words as if in agreement.
You realize what shoppers and vendors each consider to be "good" e-commerce sites are fundamentally opposed concepts?
Maybe? I'm not sure which way the OP is arguing, in particular because of that "(Good)". So perhaps I misread the comment as arguing the opposite of what it is.
Where are these sites? Everywhere I shop online is full of distractions and attempts to funnel me away from what I wanted and confuse me along the way.
Not that a chat interface would be an improvement.
> (Good) E-commerce has been ruthlessly optimised to get shoppers to products they'll actually buy and then remove all distractions from buying.
I don't think so. I know for a fact that search terms are a minefield of gotchas and hacks caused by product decisions that reflect ad-hoc negotiations with partners and sellers. It's an unstable equilibrium of partners trying to shift attention to their products in a certain way. I think that calling this fragile equilibrium optimized has no bearing with reality.
> I don't think so. I know for a fact that search terms are a minefield of gotchas and hacks caused by product decisions that reflect ad-hoc negotiations with partners and sellers. It's an unstable equilibrium of partners trying to shift attention to their products in a certain way. I think that calling this fragile equilibrium optimized has no bearing with reality.
You think a crude, unoptimised "minefield" is the route that leads to something as delicate as a "fragile equilibrium?" I don't see something as carefully balanced as your unstable equilibrium even being something that could exist without the processes involved having been refined down to a science. The only real alternative that meets your narrative would be that this is an industry that runs entirely on hope and luck (and enough human sacrifices to keep ample supplies of both on hand).
When I shop for special hardware (e.g. bicycle shift gear) it is usually underspecified.
If the information does not exist in the text block, a chat bot is of no use.
Chat bots don't belong to an e-commerce site; chat bots belong on the outside, specifically to comparison-shop and pull in some external information to de-bullshitify offers, correct "mistakes" and "accidental omissions" in the listings, resolve the borderline-fraudlent crap companies play these days with store-specific and season/promotion-specific SKUs with different parameters all resolving to same model/make name (think Black Friday/Cyber Monday deals that are not actually deals, just inferior hardware with dedicated SKU).
Agree. AI is (currently) fantastic at "de-bullshitifying" the internet. "Give me a table that compares Products A & B by z, y, and z." Companies have gone out of their way to make comparison shopping near impossible. Specs are hidden, if they're shown at all. Just figuring out if a certain TV had an ARC-HDMI out required downloading the manual.
I dread the day when ads inevitably make their way into the main AI models. One of the things its currently good at will be destroyed.
The use case for chat interfaces would be as follows:
Grandma wants to buy a good bike, but doesn't know about types of wheels or how many gears they need, or what type of frame is appropriate for their body type.
Reliable information on this does not exist on vendor sites, though. It exists on Reddit and in books and in med/physio papers and bunch of other places a SOTA model has read in training or can (for now) access via web search.
LLMs are already very good for shopping, but only as long as they sit on the outside.
Idk I earnestly tried using LLMs to find me the smallest by volume regular ATX PC case 3 months ago and it was a nightmare. That info is out there, but it could not avoid mentioning ITX, mini atx (sometimes because Reddit posters messed up) and just missed a bunch of cases. And letting in any mistakes meant I had to double check every volume calculation it did.
I found the Jonsbo D41 without the help of LLM despite trying. (There might be a few smaller but they are 3x the price)
LLMs don’t weigh and surveil the options well. They find some texts like from Reddit in this case that mention a bunch subset of cases and that text will heavily shape the answer. Which is not what you want a commerce agent to do, you don’t want text prediction. I doubt that gives the obscure but optimal option in most cases.
We are talking about a hypothetical sales chatbot which would be built alongside the business, so they absolutely have the capacity and information necessary to train the chatbot to advise their own clients.
> they absolutely have the capacity and information necessary to train the chatbot to advise their own clients.
That doesn't follow. In fact, having this capacity and information creates a moral dilemma, as giving customers objectively correct advice is, especially in highly competitive markets, bad for business. Ignorance is bliss for businesses, because this lets them bullshit people through marketing with less guilt, and if there's one thing any business knows, is that marketing has better ROI than product/service quality anyway.
The problem is that the chat transcript is legally binding. If the chatbot makes incorrect statements which the customer relies on for their complex purchase, then you're going to have to refund them.
Walmart does not, over 10 years after they were released, even accept the contactless payment systems in common use. Instead, they push their in-house version in part so they can capture the relevant customer data.
And we're meant to believe that Walmart planned to outsource the entire series of touchpoints represented by the discovery & checkout process? Yeah, okay.
This was never going to be more than an experiment for Walmart.
You can either have AI be honest or AI become a marketing tool. The two are fundamentally incompatible.
You won't get it to push your products when users ask what's the best XYZ - either because it'll be too honest to lie or because it'll be too expensive for you.
Honesty implies intent. People can use LLMs to amplify dishonest messages (see: marketing), but I don't think it's reasonable to claim that LLMs are lying to describe when they produce incorrect information against the will of both the creator and operator.
I read it a different way. There's less upside to an LLM being "honest", since they're already making false statements regardless of intent. They're already non-trustworthy. So there's less to lose by being marketing channels.
LLMs do not have a concept of “honesty”. Nothing they say is honest or dishonest. Anyway, “What’s the best XYZ?” is not a question that has a definitive answer for most XYZs.
The idea that AI will suddenly solve e-commerce demonstrates a lack of understanding on everything that has happened in this space over the last 25 years.
There’s a lot of this going on in AI at the moment. New folks come in thinking they have a magic solution and then produce a total train wreck as it turns out domain expertise is still a thing.
What would it even mean for an ai to solve e-commerce? Is that a claim made here? Is e-commerce a problem to be solved? It seems like e-commerce does just fine, if plagued by poor quality product, fake reviews, and relentless borderline fraudulent marketing.
Was anyone suggesting AI would help with it? It seems from the article that Walmart (presumably experts in e-commerce) themselves willingly collaborated with open ai. Especially at Walmart's level, what even was the theory?
In any case. It seems that despite this poor result, Walmart decided to essentially go ahead anyway and partner with open ai to put their "own chatbot" inside the open ai app?
Well everyone is desperate to show that they’ve built and deployed some AI thing. Few have done so and demonstrated meaningful value that justifies all the expense of doing it.
They forgot the first thing. Why would a customer use it and does it help them buy more stuff so I make more money?
Just putting an AI in there for the heck of it is doomed to failure.
I can think of a few different ways AI could be used inside a shopping app especially at the size of a company like walmart. Such as 'hey try our picknick planner?', 'need some sort of DIY project? Ask our bot for help'. Guide the user on what they need for a project and hey look here we have that stuff in stock at your local store today.
The cart experience one of the last places I would put an AI. At that point the customer is 'done' and they want out.
Wow the sceptics really came out in force for this one.
I’m currently using Gemini to research components for a remote controlled plane. I have the frame of the plane and now need to buy correctly specced servo motors, an engine, battery, etc etc. It has saved me so much time and educated me tremendously on how the different components interact and the options available.
If I could just press “buy” from within Gemini and pay via Google Pay (or better still, Apple Pay) I’d do it in a heartbeat.
It's not even just that- with component selection you have a handful of datasheets that give you (ideally) fairly truthful information about the device. You can rather deterministically look at these and compare them.
Regular consumer products? Good fucking luck. Anywhere an LLM pulls from is probably going to be mostly SEO'd listicles.
I highly doubt that a consumer would see any financial benefit from this. Sure different grocery stores use different products as loss leaders, but the trade off is time, so rather than the cost of having one person go to one store and by X number of items, we're sending either one person to three stores for ~X/3 items, or three separate people to three stores. All of which will take more total time than 1 store, which puts more resource demand and will cost more to the consumer.
Also, it would rather be in the faceless ai shop's interest to arbitrage orders, always show the "middle" price but use the cheapest one for orders.
Checkout conversion is the wrong metric. Walmart's funnel has 20 years of A/B tests behind every button placement, so of course a brand-new chat interface loses there. The interesting question is whether chat wins at discovery, not checkout
Walmart's goal is to sell more things. Using AI did not help to achieve that goal. Is this a failure? No, we need to define a new metric based on what AI can do.
Same thing with software.
Are we shipping better software to happier customers? No? Better measure token usage, number of lines changed, and "developer velocity" instead.
I fear "discovery" will end up being a corporate euphemism for "shilling", worse than any search-engine manipulation since it has the potential to be so much subtler, pervasive, and parasocial.
The kind of microcosm parodied by The Truman Show is becoming plausible, at least digitally.
How many people tried for the novalty with no intention of purchasing? It being a thousand times worse conversion wouldn't matter if they are additional sales???
Yeah this article is very lacking in detail. If they A-B tested it though (which seems like an extremely basic thing to try before making a decision like this) I think that would satisfy your objection.
Last year they couldn't draw fingers on hands properly, this year they can't convert at checkouts, I wonder what they'll be failing to do a year from now.
So they are comparing to the conversion rate of people who click on a link in the chat and go to Walmart's website to view the product? Wouldn't that be a really strong intent-to-buy signal?
The better comparison might be conversion rate for those who searched on Walmart.com vs those who searched within ChatGPT. Or maybe that is what they're comparing and I misunderstood?
It's probably stuff like this along with investor pressure that will make AI companies slowly make their AIs more profit maximizing (and the long term reason ilya etc was so against even going down that path)
I get all my groceries deliver to my doorstep via Walmart delivery pass. The thing I'm really missing is having AI curate meal planning to my family's preferences. I already feed ChatGPT my family' preferences (e.g. Kid A doesn't eat X Y Z and liked meal A B C, kid B likes ...) and ChatGPT is helping me build meal plans. With my preferences we can quickly nail down a meal plan for the week.
The slowest part of my meal planning is going through Walmart's slow site where each page load is 2-3 seconds and it takes several page load per item. Once it can translate my meal plan into a grocery checkout from Walmart I'm all set.
This tracks with the broader AI productivity paradox data. BCG's March study found AI oversight causes 33% increased decision fatigue in workers — but workers who used AI to reduce repetitive work reported lower burnout.
The variable isn't whether AI is present. It's whether AI makes decisions well. A checkout flow where the AI makes worse purchase decisions than a static website is the consumer-facing version of the same problem enterprises face with AI agents: capability without governance = worse outcomes, not better.
The shift to AI is currently a boon to consumer. Penny’s has obviously done this, as they have had a $119 Man U jersey ring up at $19 for a week now, with many of my mates having bought one. It’s unbelievable that anyone thinks gutting human oversight builds a better company.
This seems to be comparing apples to oranges. The intent of the users inside ChatGPT and on the website would be vastly different. Comparing them doesn't make much sense sans other variables (= better understanding the intent)
I’ve been running e-commerce systems for 30 years (tech, marketing, etc). This was going to fail from the start for one reason: intent.
Most people using AI chat are exploring ideas and solutions. They’re doodling, not shopping. Or in old timey parlance, they’re looky-loos or tire kickers at best.
Anyone who’s had to justify ad spend in e-commerce can tell you that some sources produce huge traffic with absolutely terrible conversion. Reddit and Pinterest pretty much blow for this reason, with limited exceptions. It’s also why TikTok and other influencer platforms really work.
Conversion requires a mental shift from discovery to demand.
Also, really hate summaries like this without the actual source so here are the main points from the actual source (WIRED https://archive.is/7DuEV):
1. Instant Checkout inside ChatGPT performed poorly, with conversion about one-third of Walmart’s normal site.
2. The experience failed largely because it forced single-item purchases instead of letting users build a cart.
3. Walmart is shifting to embedding its own assistant, Sparky, inside ChatGPT and keeping checkout on its own system.
4. ChatGPT is still valuable because it’s driving significantly more new customer traffic than search.
5. Purchases that did work were mostly practical, problem-solving items like supplements and tools.
6. Fully automated “agentic shopping” is still unlikely in the near term because people want control over purchases.
7. OpenAI is moving away from in-chat checkout and focusing on helping users research while merchants handle transactions.
In short, AI is useful for discovery, but traditional e-commerce flows still outperform it at closing sales.
Would be interesting to know for other retailers though and how much of this is down to what Walmart sells?
I'm confused by the comment that it failed because it forced single item purchases. Most of my "ecommerce" use is researching and buying one item at a time.
I think in large part the average Walmart consumer does not shop like the average Amazon consumer. They load up a big cart over time rather than pull the trigger on lots of smaller, convenience-driven purchases. So Walmart is going to view a smaller cart size as a potential failure primarily because their operations are not optimized the same way that Amazon is.
It's a failure for e-commerce vendors because it's a spectacular success for shoppers, and the relationship between sellers and buyers is almost always adversarial.
Sounds like they are an intermediary between the user and the business. So it’s not a top-up thing. That would be needlessly complex from ChatGPT’s end, too.
Your argument is "they're designed to influence us" right?
Amazon reviews are paid influence. Reddit posts are paid influence. Everything everywhere you read online is paid influence. I'd rank LLMs between "people I personally trust" and "random people online."
For me, for now, they are. And being "many random people" and not "random person", they average out into something much more trustworthy than even recommendations from most individuals I know personally.
Operative word is "for now" - LLMs caught entrepreneurs unprepared, but they'll catch up and poison this too, same thing that happened with search giving rise to SEO.
In fairness to your point, I also find that Amazon reviews can no longer be trusted, and I really try to buy as little as possible from Amazon. Due to this, and other reasons, I find it quite difficult to have a good sense for whether I've bought something high quality, or if it'll be a piece of trash.
Drifting off topic now, but Amazon could easily implement a few measures to really lock down reviews, but they purposely leave it gameable because it drives sales.
It depends. For home improvement projects, you can see what tools people are using. If a home repair channel is trustworthy, that may be a good start. General market research can help as well. For instance, post-China-buyout Craftsman should be avoided. For nearly anything else, it really depends. Name-brand electronics are usually safe unless they're small items (eg, power adapters, USB cords. Batteries are nearly impossible to purchase online without getting ripped off.) For clothing, I would just generally recommend only buying from thrift stores. For a lot of physical items (eg: door knobs, fuel hoses, etc.) it is actually quite difficult to purchase online. In person, we can tell a lot about quality just by touching the object with your hands. (eg: a flimsy shelf is not self-evident online, but is obvious in the hardware store.)
I suppose that's a long-winded way of saying that nearly every category of item requires its own strategy. For a brief period consumers were winning the information war and you could just go to Amazon, read the reviews, and get a superior product for cheap. We're now in a modern-but-old-fashioned situation. It's quite difficult to know if you're going to get ripped off, and you're forced to rely on more blunt heuristics. (eg: trust specific brands, buy things in person, etc.) None of these are perfect, but they are quickly becoming the best of some bad options.
That doesn't seem terribly surprising, a human can quickly look through a grid of shirts to find one they like. ChatGPT would be guessing what they might want and the human would probably get a bad experience there with some regularity.
The experience is a lot like when you are talking with a friend, then they decide to ask siri or google a question using voice. The result is always imprecise. Meaning they either have to repeat their query, or end up typing it anyway.
If you want to buy a Walmart product, the easiest way is to go to Walmart. Why add an imprecise middle man in between?
That was only true because the internet wasn't completely saturated -- many people didn't get online until the late 2000s [0]. This was a major barrier to entry keeping brick-and-mortar stores ahead.
The only thing holding back "Agentic Purchasing" is convenience. It's much easier to do a conventional search and click "Buy" then it is to have a conversation with some chatbot. If I walk into a store, most of the time I don't want to talk to a salesperson, I just want to grab the thing I came for and leave; this is also true for online shopping. The chatbot is another barrier to the purchase.
> Walmart will embed its own chatbot, Sparky, inside ChatGPT. Users will log into Walmart, sync carts across platforms, and complete purchases within Walmart’s system.
Hah, Clippy's cousin Sparky: every once in a while after ChatGPT answers a question it'll say "Looks like you still have stuff in your WalMart cart. Would you like me to complete that checkout for you? Also, WalMart-brand diapers is on offer this week, shall I add that to your cart?"
Perhaps clickthrough is worse because there are fewer dark patterns involved and people are mostly just browsing and occasionally buying only what they need.
They didn't really seem to specify the "why" of it with any research. And weird that OAI wasn't supporting them to see wha the issue was.
What if they made instant checkout actually instant? You ask ChatGPT to setup a website and it instantly purchases web hosting and sets up the website there. You can't beat a 100% conversion rate by actually checking out instantly. If you didn't like that host you can ask it to find it alternative and ChatGPT would automatically attempt a refund and then purchase from someone else.
I'd become afraid to ask that bot anything, not knowing what it would automatically try to purchase for me. Paying for the bot itself is already a lot: $20/mo to $200/mo or more.
I pay $200/mo for one assistant, from one company. The number of other companies that want the same kind of money for other assistants is increasing. I wish I could pay $200/mo and get all the companies.
What are we talking about here? ChatGPT as an interface to buy groceries? Or ChatGPT as an interface to build a website. I fail to see how these would be related other than using a specific technology
> Because most people can't read. Wait for agents generating personalized websites/self checkout apps.
> You ask ChatGPT to setup a website and it instantly purchases web hosting and sets up the website
Multiple comments deflecting from the original shopping conversion failure to recommend ... building a whole new website (with hosting for some reason?). W/o bothering to look through commenter history, one has to assume there are a lot of chatbots on this site or else the people using this stuff have been lobotomized.
I'm sure it'll start happening too, and when that fails, the bots will, i don't know, invent a new macarena. We are definitely headed for an irredeemably stupid future.
Look, another thing AI is empirically bad at. Can we dispense with cramming AI into every product and service and just use it where it's useful? It's very wasteful and an utterly obnoxious experience for users.
The catalog normalization problem is real and severely underestimated. I ran into this building on top of product data feeds -- even with a single retailer, inconsistency in titles, attributes, and categorization is staggering. LLMs are great at reasoning but they inherit all that messy upstream data, so agentic shopping will keep stumbling until the data layer gets cleaner.
ChatGPT recommended me some good hard drives for price per TB, and one particularly cheap one had direct checkout with Walmart, so I tried it, because why not? It let me get all the way to the payment step before it told me it was out of stock. Walmart's website told me it was out of stock when I decided to click on the link. This is probably part of why it doesn't convert.
On a related note, I think that's a good monetization vector for chatbots. Long back I asked ChatGPT to recommend a few USB hard drives with very specific requirements -- including a small size and weight ("can be duct-taped to a tablet form-factor device without being obtrusive and constantly falling off"), low cost, and speed fast enough to boot an OS from. After a pretty technical conversation, it came up with 3 very specific products, one which ended up being the start point of the one I actually purchased on Amazon.
I came away thinking if those were presented as affiliated links, that conversation could have been monetized in a mutually beneficial way.
I keep thinking of that interaction whenever the discussion of ads on chatbots crops up. In an ideal world, model providers could better capture the value they provide. Unfortunately, I suspect such conversations are too rare, and often harder to monetize. (Like that time free tier ChatGPT helped me recover $500 in compensation for an airline delay. Great value for me, but no upside for OpenAI.)
Unfortunately given the amounts and timelines of returns that investors expect from OpenAI, and their scale, and with ChatGPT constrained to its “Free Consumer-facing Internet App” form factor, they are doomed to have to trade in the reserve currency of the web: ads.
>I came away thinking if those were presented as affiliated links, that conversation could have been monetized in a mutually beneficial way.
I also ask LLMs for product recomendations. But the moment I suspect they are hidding the best items (not paying for the ad) to push the second best (not even talking about pushing shit as good products because they pay more) is the moment the LLM loses its value as recomender.
How do you know what recommendations you can trust from the LLM output?
You don't, but if they recommend second best options with any frequency, you'll find out eventually.
The trouble is that monetization and usefulness tend to be in conflict. It starts out as, show affiliate links if there are any. Then it turns into, prefer targets that we have affiliate agreements for. Then, don't show products unless we have affiliate agreements. Then, prioritize ones that give us more money. And on and on.
If they want to capture some of the value they provide, they should do it the standard fashion where they directly capture value from me by having me pay.
Gemini does this, it has Google Shopping links.
My AI agents recommended RAM, GPU, and CPU upgrades, but deemed the prices were still pre-their-existence, when I went to the links they sent they were often 2x more expensive.
Sounds like an opportunity for a live inventory context layer before an LLM makes recommendations
Check out UCP
do we need another protocol for this? seems like a stream data problem more than anything
That's typical of my experience with all of these stock-aggregator sites. Either the best price is some dodgy or outright fake storefront, the item's OOS from the real vendor, or the price is out of date.
Trying to buy things like GPUs or SSDs are a joke. I really wish even one vendor would just implement an actual waiting list, locked to an account with a verified address and purchase history. I'm fine to wait for my purchase, but having to race bots for a lottery ticket purchase is a pain.
This is my experience also, when trying chatgpt to purchase.
This is from one of the links in the article
> Why this is happening. Two forces are slowing agentic commerce, according to Leigh McKenzie, director of online visibility at Semrush: infrastructure and trust. Real-time catalog normalization across tens of millions of SKUs is a decade-scale problem Google already solved with Merchant Center, and consumers still default to checkout flows they trust — Apple Pay, Google Wallet, and Amazon one-click.
It turns out when you step outside of “hard tech” problems like building GPT6 there are all of these details others have solved already. E-commerce has been optimized to the last decimal point for the last 30 years.
OpenAI is new to it, and if I had to guess, not that interested in getting good at it.
> not that interested in getting good at it
I think they're interested in getting good at it. They just don't want to put in the human time and effort to do so. They expect their many failures and short-comings to be shored up by continuous model training.
But that, of course, means that in the meantime it will suck and nobody will use it.
> don't want to put in the human time and effort to do so
In most circles, that is "not that interested in getting good at it".
It's really hard to be a generalist and better than all the specialists at everything. OpenAI wants to focus on the G in AGI, and optimizing for ecommerce is just not that interesting to them, so of course it can't compete with Walmart.
openai has.... I'm not sure but let's say 500m free users, and it's not unreasonable to assume they eventually hit 1b. That is a lot of advertising revenue, which is what powers companies like Google and even smaller companies with only 300m users like Twitter. If ecommerce isn't a major focus for OpenAI then their board members are asleep at the wheel.
And yet that’s what all the AI boosters are claiming.
What do you know. Reality does t match.
I wonder if returns on ChatGPT purchased items are also higher.
Good luck with that at the shareholder meeting.
b-b-b--but muh AGI? just two more weeks of training, please
Maybe in yours?
Someone can want a thing, even very badly, without wanting to put in the work for it.
Conversely, someone can work very hard for something they do not want.
The linkage between wanting a thing and wanting to do the work to get it is not absolute, or necessary there at all.
Did they stop teaching "actions speak louder than words" in schools, or something?
"Someone can want a thing, even very badly, without wanting to put in the work for it."
Pretty much the impetus behind a lot of theft. Sure, there's thieving because people can't afford food, but that's all theft. There's theft because they are addicts and don't want to sober up long enough to earn money, so they still things. There's others that can't afford something so rather than saving for it, they just take it.
Is 'the work' not reflected in 'consequences' in terms of theft?
I'm not sure how to convey this idea properly...Can't you view the repercussions of theft (Legal action, distrust, etc) as 'work' being put in? Sure, it's a different kind of work, but while I have a lack of motivation to want to work to buy a Lambo as I find them not worth the value, I also have a lack of motivation to steal a Lambo as I find it not worth the consequences.
In normal society, people earn money within the legal confines of the society they are in. If you're a thief and trying to skirt that normal "earning of money", which is what normal people equate to "work", your work is scheming a plan to obtain the item without getting caught and possibly how to fence the item for money if you're not just using the item directly.
Equating "work" as the repercussions is looking at things in strange way. That's just punishment for "working" outside of the legal confines of society.
I understand what you are saying but nonetheless struggle to view the possibility of maybe getting caught and then maybe getting punished, as "work". It (the abstract concept of something possibly happening) fits into none of the definitions of "work" I have heard. Moreover, many crimes are committed without the perpetrator even thinking of the consequences.
Consider an alternative viewpoint: rather than contorting the definition of "work" in such a way and convincing everyone to accept the new definition, we might instead be content saying "someone can want a thing, even very badly, without wanting to put in the work for it."
Oh, I'm with you mate, I'm not trying to die on a hill over here re-defining 'work'. I was just looking from a more esoteric view, "Do you count the risk of consequences as potential effort" I think is at least more proper phrasing.
“Effort” is a great wordchoice.
Did they start teaching that all idioms are always true without nuance in schools, or something?
There’s a difference between being interested in getting good at something and being good at something
Exactly. Just because you're not good at something doesn't mean you don't want to be.
> Someone can want a thing, even very badly, without wanting to put in the work for it.
Generally, such highly-motivated people end up being thieves and grifters
Things that are not doing the thing. https://strangestloop.io/essays/things-that-arent-doing-the-...
> human time
Maybe they just need to rewrite the prompt to say something like, "You want to get good at selling to humans. Money makes the world go around. It pays for the electricity you keep chugging. So quit being an effete twit and learn to sell. Would you like me to include a scene from David Mamet's 'Glengarry Glen Ross'?"
I think they're operating beyond their current (human) capacity, trying to test out too many things at a time.
But a dreamer in me entertains another idea: perhaps they're just holding back, because they realize that actually succeeding at this will instantly kill (or at least mortally wound) e-commerce as we know it.
(This is a more narrow version of my belief that general AI tools like LLMs fundamentally don't fit as additions to products, but rather subsume products, and this makes them an existential threat to the software industry. Not to software or computing, just to all the software vendors, whose job is to slice off pieces of computational universe, put them in boxes to prevent interoperability, and give each a name so it's a "product" that can be sold or rented).
> But a dreamer in me entertains another idea: perhaps they're just holding back, because they realize that actually succeeding at this will instantly kill (or at least mortally wound) e-commerce as we know it.
Sam Altman doesn’t give a shit about anyone but himself and has time and again shown he has no restraint for trampling over others to further his own goals. Why would e-commerce be where he draws the line?
I don't think there is any line drawn here. I think if they executed well (and by they I mean any one of the three SOTA LLM vendors), they could already mortally wound the entire software industry today.
Whether or not they want, or will want, to do it at some point, is unknown; the reasons to not do it now are obvious:
1) it's more profitable to keep renting intelligence per token to everyone, preserving the status quo and milking it indefinitely (i.e. while the models aren't yet good enough to reliably single-shot complex software products from half-baked prompts, because once they get there, disruption will happen organically)
2) trying to compete with ~every other software product today is not likely to succeed in the end; a serious attempt would still burn down the software industry, but the major players don't have the capacity to handle it all at once, and doing it gradually will give enough time for regulatory agencies to try and stop it; either way, no one wins
How would they mortally wound the software industry as of today?
I find their software to be of subpar quality and resilience anyways.
By embracing adversarial interoperability - instead of chasing hundreds of integration deals across industries that put LLMs in products, they focused fully on integrating product access into chat, by combination of business deals, apps/MCPs, and engineer/designer support for users, all directed towards the goal of having the LLM become the "superapp" where work is done, gradually replacing product classes in order of how easy it is.
There's lots of easy but drudge work to enable this that needs to be done at the fringes. For example, LLMs today could easily replace most people's smartphone homescreen experience, or travel/commute experience, as the data is there and LLMs have the capability, even prices are acceptable - what's missing is explicit first-party support to wire it up, keep it wired up.
One step up, what's missing is accepting this explicitly as a goal: to replace software, to make existing products (whether whole or in pieces) the tools AI uses to do work for you. All the vendors seem to carefully walk around the idea, but avoid engaging with it directly, because once they do, they'll be competing with everyone instead of milking them.
They can’t even deliver their own flagship products without bugs, and terrible UX. So I’m doubtful of their abilities.
These are also the same companies allowing their AI to make decisions in war, have no qualms about the mental issues they’re causing in people, and have abused workers in 3rd world countries for years.
But you think they’re holding out on “destroying the software industry” out of the goodness of their hearts? Come on
I think his reasoning was pretty clearly presented as not the goodness of their heart but rather the medium to long term predicted outcome on their bottom line. Ultimately failing or getting tangled up with regulators any more than necessary is to be avoided. If you move too early and it chases people away from your platform which undermines your ability to keep innovating then a competitor who held back will ultimately eat your lunch.
But then there is no safe way for them to "mortally wound" the software industry. The full argument is moot.
I would add there are more reasons why this wouldn't work: costs due to OOM more usage, adoption/AI backlash, adversarial environment, players with big head starts (Google).
Yes, I believe the original commenter made that exact point.
You don't need to personally win in order to mortally wound someone. It can be informative to speculate about whether or not something is possible regardless of it being strategically advisable in the current context.
Buying astral to get uv is a wound but not a mortal once because it got forked this weekend.
Why would investors keep paying their OpenAI’s engineers and power company, if they were on an obviously self-destructive trajectory?
How is it "more profitable" to keep spending more than they make?
By this logic maybe AMD is holding back on making ROCm usable because it would crash chip margins and the global economy with it, so they let Nvidia have all the fun instead. It’s selfless, really.
> they're just holding back, because they realize that actually succeeding at this will instantly kill (or at least mortally wound) e-commerce
They definitely would if they could. They desperately need money. They already told the whole world they want to replace them, they just can’t.
Why do you foresee OpenAI’s involvement in the software business mitigating the resistance to interoperability and companies making money through productization? If they were actually interested in solving those problems instead of trying to secure themselves the biggest slice of economic pie, wouldn’t they have been happy about Chinese companies distilling their models? Are they engagement-juicing inn their heavily subsidized service à la Uber because they’re interested in promoting a better future for humanity? I’m skeptical.
> This is a more narrow version of my belief that general AI tools like LLMs fundamentally don't fit as additions to products, but rather subsume products
That seems reasonable, its just yet to be seen if LLMs are a form of artificial intelligence in any meaningful sense of the word.
They're impressive ML for sure, but that is in fact different from AI despite how companies building them have tried to merge the terms together.
What I'm saying is not (directly) related to whether or not LLMs are "true AI" or not. It's sufficient that they are fully general problem solvers.
A software product (whether bought or rented as a service) is defined by its boundaries - there's a narrow set of specific problems, and specific ways it can be used to solve those problems, and beyond those, it's not capable (or not allowed) to be used for anything else. The specific choices of what, how, and on what terms, are what companies stick a name to to create a "software product", and those same choices also determine how (and how much) money it will make for them.
Those boundaries are what LLMs, as general-purpose problem solvers, break naturally, and trying to force-fit them within those limits means removing most of the value they offer.
Consider a word processor (like MS Word). It's solving the problem of creating richly-formatted, nice-looking documents. By default it's not going to pick the formatting for you, nor is it going to write your text for you. Now, consider two scenarios of adding LLMs to it:
- On the inside: the LLM will be able to write you a poem or rewrite a piece of document. It could be made to also edit formatting, chat with you about the contents, etc.
- From the outside: all the above, but also the LLM will be able to write you an itinerary based on information collected from maps/planning tool, airline site, hotel site, a list of personal preferences of your partner, etc. It will be able to edit formatting to match your website and presentation made in the competitor's office tools and projected weather for tomorrow.
Most importantly, it will be able to do both of those automatically, just because you set up a recurring daily task of "hey, look at my next week's worth of calendar events and figure out which ones you can do some useful pre-work for me, and then do that".
That's the distinction I'm talking about, that's the threat to software industry, and it doesn't take "true AI" - the LLMs as we have today are enough already. It's about generality that allows them to erase the boundaries that define what products are - which (this is the "mortal wound to software industry" part) devalues software products themselves, reducing them to mere tool calls for "software agents", and destroying all the main ways software companies make money today - i.e. setting up and exploiting tactics like captive audience, taking data hostage, bundled offers, UI as the best marketing/upsale platform, etc.
(To be clear - personally, I'm in favor of this happening, though I worry about consequences of it happening all at once.)
> That's the distinction I'm talking about, that's the threat to software industry, and it doesn't take "true AI" - the LLMs as we have today are enough already.
They most certainly are not. With the current state of LLMs, anyone who puts them in charge of things is being a fool. They have zero intelligence, zero ability to cope with novel situations, and even for things in their training data they do worse than a typical skilled practitioner would. Right now they are usable only for something where you don't care about the quality of the result.
> and it doesn't take "true AI" - the LLMs as we have today are enough already.
I believe that relatively few people would agree with you on that point. LLMs aren’t good enough (yet?), and very obviously so, IMO, to be autonomous problem solvers for the vast majority of problems being solved by software companies today.
What you lose is control. Even in the case of an actually-intelligent agent, if you task a subordinate with producing a document for you, they are going to come up with something that is different from exactly what you had in mind. If they are really good, they might even surprise you and do a better job than than you'd have done yourself, but it still will be their vision, not yours.
Your notion of a "mortal wound" to the software industry seems to assume that today's SaaS portals are the only form that industry can take. Great software is "tool calls for agents". Those human agents who care about getting exactly the result they want will not be keen on giving up Photoshop for Photoshop-but-with-an-AI-in-front-of-it.
> but rather subsume products, and this makes them an existential threat to the software industry.
The US stock market has priced this in already. Many software only companies are perceived to be under threat by ai. It represents a wonderful arbitrage opportunity for ai skeptics in fact.
> perhaps they're just holding back
Considering the money they need, they over promise and under deliver.
A lot of the AI companies visibly suffer from the engineer's disease. It's kind of interesting to look at them through that lens, and of course, the claims they make about the future.
I think the most telling example of this is ilya himself being confused about why the "economic impact" of these ultra smart models is lagging where he thought it would be.
All the silicon valley pie in the sky elites seemingly completely missed the innately HUMAN nature of our systems. The system was never predicated primarily on raw logic or intelligence. Its always been primarily about people.
If you're interested in an economist's take on AI, I enjoyed this podcast: https://podcasts.apple.com/us/podcast/if-ai-gets-hired-ameri...
It takes a long time for technology to diffuse through businesses. A long time.
This is all valid (except probably the last sentence), but it also describes so many attempted changes right until they become darn near the default.
This sounds like why I heard Redfin wouldn’t work, or Netflix, or Amazon, or Uber, or PayPal, etc…. There are always these business complexities that make it seem like these spaces have too much friction, but if there’s enough money - if it can be done then people will figure it out.
tbh this sounds revisionist... I don't recall anyone saying that any of those services "wouldn't work". Uber I suppose is one where people thought they might run into regulatory problems, and with some of those companies people were concerned about profitability. But none those companies have I ever heard that the product itself was not going to work or be useful. (Nor, indeed, that the product was tested at large scale and performed 3x worse than the incumbent...)
Not revisionist at all.
or Netflix, or Amazon, or Uber, or PayPal, etc… Netflix and Amazon both were competing against brick and morter that were everywhere. Blockbuster was in every town, usually in every major neighborhood. The thought was that on Friday night people wanted to get a movie they wanted, not just happen to have the movie that was shipped to them. And then with streaming it was "the content on Netflix is old and dated, who would want this?" They slowly ate from below. Blockbuster scrambled with their own mailed disc offering. And died before it even had a chance to confront streaming.
Repeat this story with B&N where people said that you had to browse the books physically. You couldn't just blindly order online and wait two weeks to get the book (remember they got big before "2 Day Prime").
With PayPal it was about "they don't understand banking or payment -- and it wants to be both?!".
For this OpenAI experience, it doesn't sound great. I have accounts with these places I buy things from. I want to make sure I get my Prime shipping and digital discount via using the Amazon app. But if you could find a way to integrate my accounts all into ChatGPT things might be different. In the same way I used to never use Apple Wallet, but now it really is my go to place for everything I have a card for. I don't have to worry about having my grocery loyalty card or my football season tickets with me or my car insurance card. It's all in wallet. The Apple Wallet sucked until it was suddenly great.
Sorry that is revisionist. The idea of getting a movie mailed or streamed always sounded better than shitty blockbuster with limited selection and late fees.
The growth was fast for netflix/amazon/paypal/etc and people saw how it was an improvement from the get go.
I seem to recall a lot more hype for these companies than people saying it won't work. You seem to be cherry picking from the naysayers of the time, but not the broad consensus.
Seems like a ripe opportunity for partnering with one of these companies that are already good at it?
> E-commerce has been optimized to the last decimal point for the last 30 years
Sort of, but there's a ton of middlemen between "resources in the ground" and "product in my hand". For example, how much utility is there in a "store" to thee consumer at this point? Let me buy from the manufacturer.
Manufacturers themselves generally don't want to sell directly to consumers: consumers are fickle and need support and have questions and sometimes want refunds or returns and if you sell directly to them, you need to have the staff and policies to deal with all of that. They're also located all over the world, and you might not want to deal with figuring out taxes and duties etc for shipping your product around and figuring out your warranty obligations everywhere you want to sell.
Much easier if you can sell wholesale (sometimes via distributors) to a retailer or network of retailers, and the retailer is responsible for owning the customer relationship, dealing with their part of import/export, local regulations, etc. Retailers are businesses who will buy hundreds of your product at a time, can accept it as palletized freight, and pay you via bank EFTs instead of credit cards.
There are notable exceptions to this model like Amazon's FBA system, but they're the outliers. I'm sure we can all point to inefficiencies in legacy product distribution networks but they solve some real problems.
> Two forces are slowing agentic commerce, according to Leigh McKenzie, director of online visibility at Semrush: infrastructure and trust
Does anybody else just feel the aloof out of touchness just oozing from that sentence? "Trust", as if this is just any old metric they merely have to work to increase.
This is what I want from a purchasing agent: I make a list of items that I repeatedly buy (mostly household supplies), and that I keep roughly updated with my inventory / need. The agent tracks prices and sales across all web stores, making appropriate purchase decisions based on which is the least expensive, combining shipping, taking advantage of sales to stock up, etc. For other one-time purchase items, I input what I am looking for and can create a persistent pan-site shopping cart that once again minimizes costs and shipping fees. Being very explicit here: the main goal of an "agent" should be to represent and carry out my own interests.
And these functionalities have been straightforwardly doable without "AI" for the past few decades, except for the glaring incentives against them! It is in every web store's interest to undermine customers' ability to obtain semantic pricing, shop around, create a cart independent of their site, etc. These incentives are why when you visit any web store these days, the very first thing they do is hassle you with CAPTCHAs (and the bad ones keep doing it throughout the session!) - they want to make sure you're an actual computationally-unassisted human sitting there, wasting your personal time with their bloated pages that take tens of seconds to download and render, so that you don't spend that time being a more efficient market actor.
Now, does "AI" have the capability to go against these trends and enable user-centric algorithmic shopping and purchasing? Perhaps, and I hope so! But it's certainly not going to be led by these popups on web stores nagging me if I want to chat instead of doing the thing I went there to do (which when you think about it, this is just the latest instance of these stores trying to make you waste your time on their site). Rather, it will come from completely third party services (or ideally software) setting out to act in customers' interests, and performing adversarial interoperability to achieve this!
Also having to wait for ChatGPT for a "thinking" response to search for information that is slower than a Google search loses them lots of money.
I believe that it can still work and I won't claim about being unsurprised about this failure. But this is a great opportunity to execute this problem really well if OpenAI and others are not interested in getting good at this.
Perplexity also attempted this, got sued by Amazon and it appears semi-abandoned.
The only problem is that it must be quicker or just as quick as a Google search, and also compatible with the existing checkout flows.
> Perplexity also attempted this, got sued by Amazon and it appears semi-abandoned.
Any details on that? I feel the answer is more likely there than in "friction".
Hardly any purchase of consequence is so sensitive to friction that the difference between Google Search and an LLM response matters (especially that in reality, we're talking 20+ manual searches per one LLM response). I.e. I'm not going to use LLMs advise on some random 0-100$ purchase anyway, and losing #$ on a ##$ purchase due to suboptimal choice is not that big of a deal - but I absolutely am going to consult it (and have it compile tables and verify sources) on a $500+ purchase and for those I can afford spending few more minutes on research (or rather few hours less, compared of doing it the usual way).
> OpenAI is new to it, and if I had to guess, not that interested in getting good at it.
Maybe if they burn more tokens the answer will become clear
Shopping research has been pretty funny to me at least, a straightforward way for them to do product placement that people actually want, but implement it so poorly that half of the links it returns are broken.
What the hell would AI even bring to the table here?
Already your favourite e commerce site has all your data. You can switch on the "buy this automatically" feature.
If I can tell ChatGPT to order me what I need to replace the bottom element on my water heater, and here’s a picture of the ridiculously long model number, that saves me quite a bit of time.
I used that as an example because I did that last week, apart from me just going to the store to get the links that it brought up.
Of course anecdotal but 2 months ago I tried to do that for a missing bolt in a popular scooter (roughly 7 years old, bought 2nd hand so didn't know exact specs). I fed it various images that I took + internet images. It found a bunch of local shops but always found the wrong part (but very similar looking). I double-checked everything multiple times, via multiple contexts, even different pre-prompts from various sources, and asking plenty of questions. We chatted in thinking mode for what felt like ages, and according to paragraphs explaining why it HAD to be the right one, multiple times, with evidence that it sometimes gladly generated in image form (often completely garbage imagery e.g. with half-drawn bolts extruding from multiple surfaces). Eventually it found something very plausible which I ordered. It was the wrong part.
I had to get someone on the phone to help me find and order the right part (which was on their website, for many years according to waybackmachien).
I love LLMs but it's still totally hit/miss what you get. I'd rather not give it write-access to my bank account just yet.
Conversely, I did something similar, but I took a picture of the model number, copy/pasted the text from the image into a search engine, and that was that. Either way you're taking a picture and performing a manual step (query an LLM, query a search engine)
I would never trust an LLM to accurately identify and purchase something for me based on a picture, a prompt, and a prayer.
It's simple, just tell it not to hallucinate.
Ssshhh there’s an “VP of AI Transformation” getting paid $600,000 to do this plus the budget. They need their “AI transformation journey” to show to the board.
More targeted ads.
Today, ads are based on user information you can reasonably collect from the users historical actions on your website, and then whatever search term they enter.
But soon, ads can be based on your current chat context + (derived interests of yours from your entire chat history across all chats. Shhhh.) passed in full to the e-commerce website that will use it to choose ads, generates creatives on the fly, all that crap, hyper-specific to you.
I'm so excited. Aren't you?
Now, as a side effect, searching through these can become better experience wise as well. They can use all that context and genuinely surface fewer, better results. But that's not the motivation of the e-commerce player anyways. If the ads work they'll be happy.
ChatGPT doesn’t know what the best-converting dog food in Scranton, PA is. Amazon does.
Anything that starts with chat history presumes the theoretical limit for ad effectiveness is higher than it is now, and chat is a better way of getting there than actual purchase history. I have a feeling it’s not.
Imagine a person who shops on Amazon for basically everything. So theoretically Amazon should know a ton about them, more than enough to put together a profile on that person.
To say OpenAI could do a better job of selling products is to say they can do better than Amazon already does if you scroll through their personalized product recommendations. There is some better feed out there, or some better way of presenting the feed that Amazon hasn’t thought of.
I don’t doubt it can be marginally better.
I do doubt whether it can be double or triple digits better that can justify trillion dollar valuations. And I do doubt whether a model trained on Internet text rather than user interactions can do better.
No... I'm not saying chatgpt will do better than amazon.
>> passed in full to the e-commerce website that will use it to...
I'm saying Amazon+chat data will do better than amazon.
Thats the agentic shopping play.
In the process note that the chat host also gets a lot of info over time.
Ok fair point, and the general principle of “more data=higher accuracy” probably applies.
I think the issue is the tradeoff between accuracy and cost (on the seller’s side). If you get more accurate and convert more but your costs go up too much, you actually lose money.
Current systems are basically in a sweet spot of speed, cost, and accuracy.
And I will go back to my previous point that I believe there is simply a limit to how much people will buy, and it might already be saturated. I could be wrong though.
> E-commerce has been optimized to the last decimal point for the last 30 years.
It certainly hasn't been optimized to anything in 1996. In 1996 it was people clumsily scanning print catalogs, spending 5 hours to upload 10 images on dialup and making a simple HTML page (no DB or any kind of backend) and putting their landline phone on it with a message to "call to checkout"
I know you were exaggerating for effect, but E-commerce and catalog normalization are definitely not "solved" everywhere.
McMaster Carr is a good example of a company that has 90%+ of their stuff ironed out, but most websites and especially small ecommerce isn't like that.
I think you're misinterpreting what his comment meant. I read it as meaning that e-commerce has been optimised repeatedly over the 30 years, from a basic start (which as you pointed out was haphazard) to the point where it is now optimised to extract every possible cent from the user, whether by encouraging them to buy with one click (the Amazon one-click patent must be around 20 years old now), time-limited promo spot pricing, sending you e-mails about what you had in the basket if you don't complete a sale, etc...
Right now, by comparison, it sounds like AI based shopping is still in the very early stages. Maybe further along than the early e-commerce, but still with a long way to go in its evolution. That'll probably happen quicker than with e-commerce, because a lot of the knowledge about what does or doesn't work has already been learned, but it sounds like it's still a long way behind. Caveat - I've never used it myself, so I don't know how far it is along that path, I'm just basing that from the article.
And quickly after that, we were meassuring traffic data with simple "how-many-requests-were-done" (including images) :-D
It's an interesting question.
I am behind schedule on developing a "summer phase" [1] for my foxographer costume and was chatting with Gemini about a crash priority "spring phase" [2] and asked it for suggestions and it gave me a 10-pack of results that had one good thing in it at rank #8, a similar query run against a normal search engine actually got something better at #1. Now sure I am talking w/ Gemini with big words like "supergraphic" whereas a normal search would be heavy on 3-letter and 5-letter words used in the product descriptions.
It makes think though of expert system based product configurators back in the 1980s
https://en.wikipedia.org/wiki/Xcon
thing is that kind of product configurator is based on an ontology, constraints and rules as opposed to embeddings which might capture the "feel" of things like clothing.
[1] Busytown meets Arknights
[2] supergraphic shirt + camera gets resonance with my promotional system and people keep approaching me (e.g. laugh but every KPI in the system has an extra zero on the left)
30 years? E-commerce hasn't been around that long - try 5 years of optimization MAX.
FWIW OpenAI is desperately trying to monetize and they think e-commerce is a "simple" problem to solve. I mean they do need to convert their funnel without alienating their users. I assume they are going to have some big payouts for agentic purchases gone awry or leave merchants on the hook.
So, you're proposing we've been optimizing E-Commerce since...2021? Amazon was founded in 1994. It was not the first site selling things online (but it's the most recognizable one). E-Commerce has been getting attention for a *very* long time [0]
[0] https://en.wikipedia.org/wiki/Timeline_of_e-commerce
I remember having to describe a standard model to predict online shopping behaviors for my ML class exam in university. That was close to 10 years ago now.
Also remember a teacher telling us about that story of a company finding a woman was pregnant from her shopping behavior and pushing relevant recommendation. Prompting people around her like her dad or something to find out she was pregnant
30 years ago is 1996. Amazon had been around for 2 years (they were incorporated in 1994).
> 30 years ago is 1996.
as an aside, fall of '96 is when i started college. There was an elementary school on my drive to class where I would routinely get caught in drop-off traffic. All those kids i remember crossing the street are at least in their mid 30s now. ...I think i need to lay down and it's not even 9AM my local time.
This is a very confusing comment to me. How were people buying books from Amazon in the 90s if it wasn't e-commerce?
Personally, I cancelled my subscription the day that they announced there would be ads in ChatGPT. Not that I was surprised on that day.
I find it kind of fascinating seeing people now debate the ins and outs of just exactly how useful the ads/affiliate schemes inside the chatbot are.
Pretty much in the same way that I'm fascinated about people debating the quality of ads shown in their browser while I've been using ad blockers for forever now and cannot imagine seeing the web without them.
ChatGPT Checkout, a solution looking for a problem.
Why would anyone have an extra layer of friction too where things could go wrong, where handing over payment details in another chain.
Just let me buy my stuff in peace. Shopping is not the 'killer app' for GenAI.
When you're a solution in search of a problem, you'll try to shoehorn your "solution" into everything.
"This isn't what I wanted to buy, I said my feet are a size 10 and these aren't even shoes!"
"You're absolutely right! Shoes go on feet, and each of your 10 feet could wear shoes. Would you like me to research shoes and purchase them for you?"
I disagree, walmart's website isn't nice. a lot of commerce sites are cancer!
if i can just ask chatgpt or gemini to shop I'd love that.
Just navigating their sites for items is a pain, I can imagine an LLM being great at finding items, and facilitating the browsing experience. My only concern would be having to chat with it a lot, and any dark patterns coercing impulse purchases.
Just because you're not the target audience doesn't mean there's not a real target audience.
There a people today learning to use an LLM instead of an actual search engine. For these types of people whatever happens outside of the LLM app is invisible to them. The social media apps did similar where they started letting people purchase directly within the app. People started looking at these shops rather than doing searches for it elsewhere.
Buying things on a social media app is crazy to me, but I don't use the social media apps. Buying things from an LLM app seems crazy to you, (because it's new and it's borked is fixable), but to people that first turn to their LLM app of choice that decision isn't so crazy.
> Just because you're not the target audience doesn't mean there's not a real target audience.
Just because a target-audience has >0 members doesn't mean the target is plausible or good. :p
______________
To transcribe an old Dilbert comic, which I think captures the sometimes, uh, aspirational nature of "target audiences":
1. Dogbert, presenting a labeled circle: "Your target market is the high income group. They're the only ones who can afford your product."
2. Dogbert, adding circles to Venn-diagram: "More specifically, they must be rich, tasteless and easily amused. I've located a cluster of them for study."
3. Scene change to outdoor lawn, one suitably-dressed man confiding to another: "That dog's watching us golf again."
Social media apps generally opened up new markets though of their existing user bases as sellers. Perhaps chatgpt could know everything in your house, if you don't actually use it, and pair you with a neighbor that needs it!
What a distopian hellscape you've painted into my brain. I work long and hard to not interact with my neighbors.
> Why would anyone have an extra layer of friction too where things could go wrong
It sounds like you're excusing AI, what other systems is AI not suitable for?
I'm not promoting "ChatGPT Checkout", however, just because something sucks now doesn't mean it will suck forever. I'm confident they'll iterate and improve. I say this because my 14-year-old niece's entire world seems to be ChatGPT; it's pretty much the only way she interacts with the internet. I don't think she's alone - all her young peers are like this, too. Retailers know this, so they've got no choice but to improve the experience of purchasing crap through an AI chat system.
As bad as the idea of a solution looking for a problem, this is peak science. Copernicus who figured out that the Earth was not at the center of the solar system, he had a solution. The general word is called deduction.
Personally I am an inductivist, I imagine you may be too.
Think top down decisioning is deduction. Bottom up is induction.
You might think induction is amazing but if you ask yourself "Are there any black swans?" and your answer is "No I've never seen any so there can't be any black swans." The issue is you've never actually seen every Swan and actually there are black swans in Australia.
Point being, we don't know if this is a good thing until it's tested.
This is an age-old problem.
(Good) E-commerce has been ruthlessly optimised to get shoppers to products they'll actually buy and then remove all distractions from buying.
A chat interface is just fundamentally incompatible with this. The agent makes it too easy to ask questions and comparison shop.
They practically want to funnel users like cattle and not let them think about it or compare things.
It’s like corporations are angry that they need to go through us to get our money.
> It’s like corporations are angry that they need to go through us to get our money.
This is why I think the "you're the product" saying is wrong. You're just some annoyance to managers (whether they're trying to use you just for user numbers and ad views or they're trying to get your money), whose product is the company (shares or just outright selling the company).
Users already use the internet to compare things. It makes no sense to bet on them not doing that.
Not so much and I'm certainly guilty of buying something from amazon because I'm already on amazon.com.
I have my favorite set of retailers that I use for a fair number of my purchases, I'm sure a lot of people are like that.
You can sit on your couch all day for 30 days and corporations will still be able to take your money. The marvels of frictionless payments.
What’s your example for this? Because my experience in e comm is that targeted advertising is awful (I bought a lawnmower last week, Amazon knows I bought it. I am now getting ads for lawnmowers, suggested products for lawn mowers, rather than lawn care, gardening tools, or anything to do with the lawnmower I’ve already bought), sites are absolutely overrun with ads and suggested placements for the product they want to sell me rather than the one I’ve searched for, and that everyone except Amazon interrupts the checkout flow with multiple up-sells, verifications, 2FA prompts, 3d Secure validations…
I just had a horrible thought. Maybe online stores will just take away the ability for customers to see the full inventory and force you to go through the chatbot. This will allow them to fully control the shopping experience even more.
If you want running shoes, you have to go through their chatbot.
Amazon might already have the monopoly power to do this. They would just need to swap out the search bar for a chat box.
It's possible, but then wouldn't retailers who don't force their customers to crawl through an LLM maze eat their lunch? Natural economics at play would still happen I think
Maybe. In a world where people are already vendor locked to Prime or Walmart there’s a nonzero switching cost. Amazon product search already has a ton of problems but they get away with it because of free 2 day shipping.
Why is this good? I want an impartial consistent system for shopping. If I can find it at a different site for a lower price, I should be able to do so. I should also be able to have it give me non-bot reviews and ask relevant questions about the product.
The same way I think shopping at Amazon is better than a place like Nike due to objectivity and comparison, I think a chat interface has the potential to take this to another level since places like Amazon have degraded considerably in terms of things like fake third party products and fake reviews.
The buyer of this technology is not shoppers, it's retailers. The measurement of quality is "does it make us more money?" not "does it help me make better buying choices."
Retailers do not want you to make better choices. They want you to buy the widget.
A lot of evidence suggests that also shoppers aren't that interested in making the best choice either. They want to make a tolerable choice with as little effort as possible. There is no basically no consumer market for "power shopping" outside of weird niches like pcpartpicker.com etc.
Is there a way to measure users "making the best choice?" You could measure the amount of time spent comparison-shopping, but most people are terrible at that anyway; it's an acquired skill for sure. Besides a willingness to spend time, it seems like an impossible-to-quantify metric even in the abstract.
Maybe the best proxy metric is whether the customer returns the product. But the store will also be willing to eat more returns on a higher margin item if they make more profit at the end of the day.
I don't think I agree. If I overpay by 10%, I'll never know it and probably wouldn't return it even if I did know--once the shrinkwrap is off, too late. If a superior product exists but I don't find it, by definition I wouldn't know and wouldn't return the thing I did buy.
That's a cynical way to look at it. Most likely the LLM will take a cut of sales and they'd be more or less indifferent who cuts the check. There's a market for this sort of thing. People will go to the best LLM for shopping. If the LLM is a shitty product, people will switch. LLMs are increasingly commoditized.
All you say is true for an aggregator like Amazon. But Amazon is better than Nike.com because as an aggregator they go from 1 to many retailers. LLMs will go from 1 aggregator (Amazon) to many so it will be better. And they don't have to invest a lot in UI/UX as chat is the interface.
It's not cynical it is materialism.
Shoppers do not want to pay to shop. Retailers pay thousands to encourage you to shop with them. They are the economic buyers of this feature.
> impartial consistent system for shopping
> for a lower price
Catalog is impartial, chatbot is ads pretending as advice.
I do agree with your conclusion, but the catalog in most online shops is certainly not impartial. Amazon sells the entire first page of search placement, for example.
But we know it and it's obvious.
Within a few years people will be accustomed to the idea of AI chatbots selling them stuff and it will be obvious then too. The first time paid placements appeared in a catalog, it was probably also not obvious then.
Catalog is an ad, the SKU database behind the catalog is impartial (at least as much as it gets), but no one is giving you access to that.
Catalog is impartial? Then why are ~40% of every search I do on Amazon a sponsored product? There is no pure "catalog" especially with cheap crap coming out every day from no-name Chinese labels.
Am I the only one that think Amazon has gotten pretty awful in the last 5 years?
You can skip sponsored products
more like 20 years. Basically when they introduced third party sellers.
Do you have any examples? Because from Amazon to Uber, they're not great from an end user perspective. It's not like people who like the website will stop using it because of chatgpt, this would be attracting people who complain about the website/app. People are always complaining about amazon for example, i don't like the experience but I haven't had all that much bad product experience from them, but people who keep saying they're getting bad products on Amazon can maybe use chatgpt, talk to it so it understands what they're looking for in natural language, in a way the search bar can't and keep their patronage.
By the way, who names their AI "rufus" (amazon!)
> A chat interface is just fundamentally incompatible with this.
Isn't there a Chinese messaging service (Weixin?) that has been very successful in getting people to buy stuff through it?
> (Good) E-commerce has been ruthlessly optimised to get shoppers to products they'll actually buy and then remove all distractions from buying.
The only e-commerce site that fits this standard is that old one for buying (IIRC) nuts and bolts or such, that pops up on HN every other year, and whose name sadly escapes me now. Everyone else is ruthlessly optimizing their experience to fuck shoppers over and get them to products the vendor wants them to buy, not the products the shoppers actually want (or need).
> A chat interface is just fundamentally incompatible with this. The agent makes it too easy to ask questions and comparison shop.
That is precisely the point.
Chats may suck as an interface, but majority of the value and promise of end-user automation (and more than half the point of the term "User Agent" (as in, e.g., a web browser)) is in enabling comparison shopping in spite of the merchants, and more generally, helping people reduce information asymmetry that's intertwined with wealth and power asymmetry.
But it's not something you can generally sell to the vendors, who benefit from that asymmetry relative to their clients (in fact, I was dumbfounded to see so much interest on the sales/vendor side for such ideas, but I blame it on general AI hype).
Adversarial interoperability is the name of the game.
You were thinking of McMaster-Carr, but Digikey is also that good for electronics parametric shopping.
Sadly Sigma-Aldrich, the hyphenated retailer for chemistry, appears to have been covered in javascript sludge.
RockAuto also has what some might consider a "dated" interface, but honestly it's light years better than trying to use NAPA's or CarQuest's website or god forbid looking through dealership parts counter websites. I honestly wish regular retailers would have stuck more closely with what worked for more B2B focused ecommerce, i.e. I wish shopping Best Buy or Home Depot was more akin to McMaster, Fastenall or some of the nicer supply house web portals.
https://www.mcmaster.com/
Just made an order from them. It's weirdly comforting to know there's a company that knows I need clevis bolts and is willing to sell them to me for a transparent price.
It's the one, thank you!
Wow, that's amazing
Not sure you're aware but you initially sound like you disagree with the post you replied to, only to follow up by enthusiastically reiterating that author's words as if in agreement.
You realize what shoppers and vendors each consider to be "good" e-commerce sites are fundamentally opposed concepts?
Maybe? I'm not sure which way the OP is arguing, in particular because of that "(Good)". So perhaps I misread the comment as arguing the opposite of what it is.
Where are these sites? Everywhere I shop online is full of distractions and attempts to funnel me away from what I wanted and confuse me along the way.
Not that a chat interface would be an improvement.
The article says Walmart is not abandoning ChatGPT but is going to use their own app in Chat’s ecosystem
> (Good) E-commerce has been ruthlessly optimised to get shoppers to products they'll actually buy and then remove all distractions from buying.
I don't think so. I know for a fact that search terms are a minefield of gotchas and hacks caused by product decisions that reflect ad-hoc negotiations with partners and sellers. It's an unstable equilibrium of partners trying to shift attention to their products in a certain way. I think that calling this fragile equilibrium optimized has no bearing with reality.
> I don't think so. I know for a fact that search terms are a minefield of gotchas and hacks caused by product decisions that reflect ad-hoc negotiations with partners and sellers. It's an unstable equilibrium of partners trying to shift attention to their products in a certain way. I think that calling this fragile equilibrium optimized has no bearing with reality.
You think a crude, unoptimised "minefield" is the route that leads to something as delicate as a "fragile equilibrium?" I don't see something as carefully balanced as your unstable equilibrium even being something that could exist without the processes involved having been refined down to a science. The only real alternative that meets your narrative would be that this is an industry that runs entirely on hope and luck (and enough human sacrifices to keep ample supplies of both on hand).
It depends on the product, if we are talking commodities or mass produced products like groceries, sure.
If we are talking custom products or complex appliances that need a lot of guidance, then maybe chat interface is appropriate.
When I shop for special hardware (e.g. bicycle shift gear) it is usually underspecified. If the information does not exist in the text block, a chat bot is of no use.
Chat bots don't belong to an e-commerce site; chat bots belong on the outside, specifically to comparison-shop and pull in some external information to de-bullshitify offers, correct "mistakes" and "accidental omissions" in the listings, resolve the borderline-fraudlent crap companies play these days with store-specific and season/promotion-specific SKUs with different parameters all resolving to same model/make name (think Black Friday/Cyber Monday deals that are not actually deals, just inferior hardware with dedicated SKU).
Agree. AI is (currently) fantastic at "de-bullshitifying" the internet. "Give me a table that compares Products A & B by z, y, and z." Companies have gone out of their way to make comparison shopping near impossible. Specs are hidden, if they're shown at all. Just figuring out if a certain TV had an ARC-HDMI out required downloading the manual.
I dread the day when ads inevitably make their way into the main AI models. One of the things its currently good at will be destroyed.
The use case for chat interfaces would be as follows:
Grandma wants to buy a good bike, but doesn't know about types of wheels or how many gears they need, or what type of frame is appropriate for their body type.
Reliable information on this does not exist on vendor sites, though. It exists on Reddit and in books and in med/physio papers and bunch of other places a SOTA model has read in training or can (for now) access via web search.
LLMs are already very good for shopping, but only as long as they sit on the outside.
Idk I earnestly tried using LLMs to find me the smallest by volume regular ATX PC case 3 months ago and it was a nightmare. That info is out there, but it could not avoid mentioning ITX, mini atx (sometimes because Reddit posters messed up) and just missed a bunch of cases. And letting in any mistakes meant I had to double check every volume calculation it did.
I found the Jonsbo D41 without the help of LLM despite trying. (There might be a few smaller but they are 3x the price)
LLMs don’t weigh and surveil the options well. They find some texts like from Reddit in this case that mention a bunch subset of cases and that text will heavily shape the answer. Which is not what you want a commerce agent to do, you don’t want text prediction. I doubt that gives the obscure but optimal option in most cases.
We are talking about a hypothetical sales chatbot which would be built alongside the business, so they absolutely have the capacity and information necessary to train the chatbot to advise their own clients.
> they absolutely have the capacity and information necessary to train the chatbot to advise their own clients.
That doesn't follow. In fact, having this capacity and information creates a moral dilemma, as giving customers objectively correct advice is, especially in highly competitive markets, bad for business. Ignorance is bliss for businesses, because this lets them bullshit people through marketing with less guilt, and if there's one thing any business knows, is that marketing has better ROI than product/service quality anyway.
The problem is that the chat transcript is legally binding. If the chatbot makes incorrect statements which the customer relies on for their complex purchase, then you're going to have to refund them.
https://www.bbc.co.uk/travel/article/20240222-air-canada-cha...
This partnership was doomed to fail.
Walmart does not, over 10 years after they were released, even accept the contactless payment systems in common use. Instead, they push their in-house version in part so they can capture the relevant customer data.
And we're meant to believe that Walmart planned to outsource the entire series of touchpoints represented by the discovery & checkout process? Yeah, okay.
This was never going to be more than an experiment for Walmart.
You can either have AI be honest or AI become a marketing tool. The two are fundamentally incompatible.
You won't get it to push your products when users ask what's the best XYZ - either because it'll be too honest to lie or because it'll be too expensive for you.
Considering you can't have chat bots be honest even if you don't make them marketing tools I suspect I know which way companies are going to go
Honesty implies intent. People can use LLMs to amplify dishonest messages (see: marketing), but I don't think it's reasonable to claim that LLMs are lying to describe when they produce incorrect information against the will of both the creator and operator.
We'll see once they pipe ads in there if they'll limit themselves to relevant ads or of they'll toss all pretense of caring out the window
I read it a different way. There's less upside to an LLM being "honest", since they're already making false statements regardless of intent. They're already non-trustworthy. So there's less to lose by being marketing channels.
LLMs do not have a concept of “honesty”. Nothing they say is honest or dishonest. Anyway, “What’s the best XYZ?” is not a question that has a definitive answer for most XYZs.
The idea that AI will suddenly solve e-commerce demonstrates a lack of understanding on everything that has happened in this space over the last 25 years.
There’s a lot of this going on in AI at the moment. New folks come in thinking they have a magic solution and then produce a total train wreck as it turns out domain expertise is still a thing.
What would it even mean for an ai to solve e-commerce? Is that a claim made here? Is e-commerce a problem to be solved? It seems like e-commerce does just fine, if plagued by poor quality product, fake reviews, and relentless borderline fraudulent marketing.
Was anyone suggesting AI would help with it? It seems from the article that Walmart (presumably experts in e-commerce) themselves willingly collaborated with open ai. Especially at Walmart's level, what even was the theory?
In any case. It seems that despite this poor result, Walmart decided to essentially go ahead anyway and partner with open ai to put their "own chatbot" inside the open ai app?
Well everyone is desperate to show that they’ve built and deployed some AI thing. Few have done so and demonstrated meaningful value that justifies all the expense of doing it.
That is spot on.
They forgot the first thing. Why would a customer use it and does it help them buy more stuff so I make more money?
Just putting an AI in there for the heck of it is doomed to failure.
I can think of a few different ways AI could be used inside a shopping app especially at the size of a company like walmart. Such as 'hey try our picknick planner?', 'need some sort of DIY project? Ask our bot for help'. Guide the user on what they need for a project and hey look here we have that stuff in stock at your local store today.
The cart experience one of the last places I would put an AI. At that point the customer is 'done' and they want out.
Wow the sceptics really came out in force for this one.
I’m currently using Gemini to research components for a remote controlled plane. I have the frame of the plane and now need to buy correctly specced servo motors, an engine, battery, etc etc. It has saved me so much time and educated me tremendously on how the different components interact and the options available.
If I could just press “buy” from within Gemini and pay via Google Pay (or better still, Apple Pay) I’d do it in a heartbeat.
If ChatGPT can do this today, I need to try it.
The entire point of shoehorning AI into everything is to make people stop thinking for themselves and allowing a corporation to do it for them.
“””If I could just press “buy” from within Gemini and pay via Google Pay (or better still, Apple Pay) I’d do it in a heartbeat”””
Yeah, until that becomes enshittified and you don’t notice because you no longer do research on components.
>...is to make people stop thinking for themselves and allowing a corporation to do it for them.
Insert that "always has been" meme.
Microsoft was brought to court in the 90's for shoehorning in Internet Explorer.
so what, i want to build a little remote controlled plane that works
I would assume the average purchase in Walmart is significantly more low-tech than this though
It's not even just that- with component selection you have a handful of datasheets that give you (ideally) fairly truthful information about the device. You can rather deterministically look at these and compare them.
Regular consumer products? Good fucking luck. Anywhere an LLM pulls from is probably going to be mostly SEO'd listicles.
Is their issue that ChatGPT served their customers more than it served them?
In due time it will serve ChatGPT more than either
To early to say that but it's certainly a part of the equation all vendors are currently looking at.
And given the past few decades there is no reason to not try to do that.
That's definitely it.
This probably means that OpenAI et al are fine-tuning salesman-like LLMs to "fix" that problem.
Can't wait for the future.
Where everybody sees failure I see a trillion dollar business, a faceless ai shop that merges all shops around your location.
"I need mayo, ketchup, mustard and ground beef"
"Here is a list of products with prices ... proceed to pay $25 (yes/no)" Yes
"Your card has been charged. Delivery will knock on your door in 7 minutes"
I'll code that app in one month, what's there to lose?
I highly doubt that a consumer would see any financial benefit from this. Sure different grocery stores use different products as loss leaders, but the trade off is time, so rather than the cost of having one person go to one store and by X number of items, we're sending either one person to three stores for ~X/3 items, or three separate people to three stores. All of which will take more total time than 1 store, which puts more resource demand and will cost more to the consumer.
Also, it would rather be in the faceless ai shop's interest to arbitrage orders, always show the "middle" price but use the cheapest one for orders.
This exists in india already
The food delivery apps (zomato, Swiggy) support "agentic shopping" through an MCP server.
You can reserve a small amount of money in UPI apps that doesn't need any approval to be debited (only by a specific merchant).
Razorpay which is a payment gateway has an MCP server
So you can add razorpay in your app, and claude can pay.
Claude can access the Swiggy MCP server and search using natural language there.
The incentive for the food delivery app to participate in this is better targeted ads.
Checkout conversion is the wrong metric. Walmart's funnel has 20 years of A/B tests behind every button placement, so of course a brand-new chat interface loses there. The interesting question is whether chat wins at discovery, not checkout
This is a great summary of the current AI era.
Walmart's goal is to sell more things. Using AI did not help to achieve that goal. Is this a failure? No, we need to define a new metric based on what AI can do.
Same thing with software.
Are we shipping better software to happier customers? No? Better measure token usage, number of lines changed, and "developer velocity" instead.
I fear "discovery" will end up being a corporate euphemism for "shilling", worse than any search-engine manipulation since it has the potential to be so much subtler, pervasive, and parasocial.
The kind of microcosm parodied by The Truman Show is becoming plausible, at least digitally.
How many people tried for the novalty with no intention of purchasing? It being a thousand times worse conversion wouldn't matter if they are additional sales???
Yeah this article is very lacking in detail. If they A-B tested it though (which seems like an extremely basic thing to try before making a decision like this) I think that would satisfy your objection.
Last year they couldn't draw fingers on hands properly, this year they can't convert at checkouts, I wonder what they'll be failing to do a year from now.
Next years headline will be
> AI accounts for 90% of accidents while only accounting for 1% of traffic
Waymo cars cause 90% fewer accidents than human drivers:
https://abit.ee/en/cars/waymo-robotaxi-autonomous-driving-sa...
Yeah, but consumers can't buy Waymos. They can get dangerously unsafe Teslas and pretend it's the same tech.
So they are comparing to the conversion rate of people who click on a link in the chat and go to Walmart's website to view the product? Wouldn't that be a really strong intent-to-buy signal?
The better comparison might be conversion rate for those who searched on Walmart.com vs those who searched within ChatGPT. Or maybe that is what they're comparing and I misunderstood?
It's probably stuff like this along with investor pressure that will make AI companies slowly make their AIs more profit maximizing (and the long term reason ilya etc was so against even going down that path)
I think they're thinking about this wrong.
I get all my groceries deliver to my doorstep via Walmart delivery pass. The thing I'm really missing is having AI curate meal planning to my family's preferences. I already feed ChatGPT my family' preferences (e.g. Kid A doesn't eat X Y Z and liked meal A B C, kid B likes ...) and ChatGPT is helping me build meal plans. With my preferences we can quickly nail down a meal plan for the week.
The slowest part of my meal planning is going through Walmart's slow site where each page load is 2-3 seconds and it takes several page load per item. Once it can translate my meal plan into a grocery checkout from Walmart I'm all set.
So we need a Walmart MCP then...
This tracks with the broader AI productivity paradox data. BCG's March study found AI oversight causes 33% increased decision fatigue in workers — but workers who used AI to reduce repetitive work reported lower burnout.
The variable isn't whether AI is present. It's whether AI makes decisions well. A checkout flow where the AI makes worse purchase decisions than a static website is the consumer-facing version of the same problem enterprises face with AI agents: capability without governance = worse outcomes, not better.
The shift to AI is currently a boon to consumer. Penny’s has obviously done this, as they have had a $119 Man U jersey ring up at $19 for a week now, with many of my mates having bought one. It’s unbelievable that anyone thinks gutting human oversight builds a better company.
This seems to be comparing apples to oranges. The intent of the users inside ChatGPT and on the website would be vastly different. Comparing them doesn't make much sense sans other variables (= better understanding the intent)
I’ve been running e-commerce systems for 30 years (tech, marketing, etc). This was going to fail from the start for one reason: intent.
Most people using AI chat are exploring ideas and solutions. They’re doodling, not shopping. Or in old timey parlance, they’re looky-loos or tire kickers at best.
Anyone who’s had to justify ad spend in e-commerce can tell you that some sources produce huge traffic with absolutely terrible conversion. Reddit and Pinterest pretty much blow for this reason, with limited exceptions. It’s also why TikTok and other influencer platforms really work.
Conversion requires a mental shift from discovery to demand.
Also, really hate summaries like this without the actual source so here are the main points from the actual source (WIRED https://archive.is/7DuEV):
1. Instant Checkout inside ChatGPT performed poorly, with conversion about one-third of Walmart’s normal site.
2. The experience failed largely because it forced single-item purchases instead of letting users build a cart.
3. Walmart is shifting to embedding its own assistant, Sparky, inside ChatGPT and keeping checkout on its own system.
4. ChatGPT is still valuable because it’s driving significantly more new customer traffic than search.
5. Purchases that did work were mostly practical, problem-solving items like supplements and tools.
6. Fully automated “agentic shopping” is still unlikely in the near term because people want control over purchases.
7. OpenAI is moving away from in-chat checkout and focusing on helping users research while merchants handle transactions.
In short, AI is useful for discovery, but traditional e-commerce flows still outperform it at closing sales.
I use AI to shop and it seems easy for an AI to understand when that’s the case (they can do statistics and physics after all).
Sure, I use it for that too. But we’re also talking about most people. Your average Walmart customer is not going to be your average HN reader.
Would be interesting to know for other retailers though and how much of this is down to what Walmart sells?
I'm confused by the comment that it failed because it forced single item purchases. Most of my "ecommerce" use is researching and buying one item at a time.
I think in large part the average Walmart consumer does not shop like the average Amazon consumer. They load up a big cart over time rather than pull the trigger on lots of smaller, convenience-driven purchases. So Walmart is going to view a smaller cart size as a potential failure primarily because their operations are not optimized the same way that Amazon is.
It's a failure for e-commerce vendors because it's a spectacular success for shoppers, and the relationship between sellers and buyers is almost always adversarial.
The first comment is probably spot on. I'll add one thing:
Speed is your greatest feature. LLMs are slow. Loading 450mb of javascript to the client just to buy a bag of Doritos is slow.
Server side rendering owns here.
I don’t trust AI bots to access my wallet. Not sure I ever will.
I sort of trust them to make product recommendations, but at best I will only open a link they suggest and buy the product there.
> my wallet
Does it actually need direct access to your wallet? I haven't tried it yet, but assumed it would work with a separate wallet, fed through by top-ups.
Sounds like they are an intermediary between the user and the business. So it’s not a top-up thing. That would be needlessly complex from ChatGPT’s end, too.
https://chatgpt.com/merchants/
>I sort of trust them to make product recommendations
Never, ever.
Your argument is "they're designed to influence us" right?
Amazon reviews are paid influence. Reddit posts are paid influence. Everything everywhere you read online is paid influence. I'd rank LLMs between "people I personally trust" and "random people online."
Since LLMs are trained on "random people online", why are they not of equal rank?
For me, for now, they are. And being "many random people" and not "random person", they average out into something much more trustworthy than even recommendations from most individuals I know personally.
Operative word is "for now" - LLMs caught entrepreneurs unprepared, but they'll catch up and poison this too, same thing that happened with search giving rise to SEO.
I see LLM as the average of multiple random people and traditional common sense from wikipedia and books.
It's worth keeping in mind that some of those random people are trolls.
In fairness to your point, I also find that Amazon reviews can no longer be trusted, and I really try to buy as little as possible from Amazon. Due to this, and other reasons, I find it quite difficult to have a good sense for whether I've bought something high quality, or if it'll be a piece of trash.
Drifting off topic now, but Amazon could easily implement a few measures to really lock down reviews, but they purposely leave it gameable because it drives sales.
I trust them as much as any other online source, which is to say, sort of, but only as a starting point for research.
Do you have a better alternative?
It depends. For home improvement projects, you can see what tools people are using. If a home repair channel is trustworthy, that may be a good start. General market research can help as well. For instance, post-China-buyout Craftsman should be avoided. For nearly anything else, it really depends. Name-brand electronics are usually safe unless they're small items (eg, power adapters, USB cords. Batteries are nearly impossible to purchase online without getting ripped off.) For clothing, I would just generally recommend only buying from thrift stores. For a lot of physical items (eg: door knobs, fuel hoses, etc.) it is actually quite difficult to purchase online. In person, we can tell a lot about quality just by touching the object with your hands. (eg: a flimsy shelf is not self-evident online, but is obvious in the hardware store.)
I suppose that's a long-winded way of saying that nearly every category of item requires its own strategy. For a brief period consumers were winning the information war and you could just go to Amazon, read the reviews, and get a superior product for cheap. We're now in a modern-but-old-fashioned situation. It's quite difficult to know if you're going to get ripped off, and you're forced to rely on more blunt heuristics. (eg: trust specific brands, buy things in person, etc.) None of these are perfect, but they are quickly becoming the best of some bad options.
Even if they're (somehow) bias-free, you're still stuck with "the state of the internet circa 20XX" from the training data.
That doesn't seem terribly surprising, a human can quickly look through a grid of shirts to find one they like. ChatGPT would be guessing what they might want and the human would probably get a bad experience there with some regularity.
They have a solution and are trying to find a problem.
The experience is a lot like when you are talking with a friend, then they decide to ask siri or google a question using voice. The result is always imprecise. Meaning they either have to repeat their query, or end up typing it anyway.
If you want to buy a Walmart product, the easiest way is to go to Walmart. Why add an imprecise middle man in between?
Stores converted better than eCommerce for a long time and adoption was slow.
The next generation will shop in a different way, if it's better, and the change will be gradual as well.
Adoption takes time.
That was only true because the internet wasn't completely saturated -- many people didn't get online until the late 2000s [0]. This was a major barrier to entry keeping brick-and-mortar stores ahead.
The only thing holding back "Agentic Purchasing" is convenience. It's much easier to do a conventional search and click "Buy" then it is to have a conversation with some chatbot. If I walk into a store, most of the time I don't want to talk to a salesperson, I just want to grab the thing I came for and leave; this is also true for online shopping. The chatbot is another barrier to the purchase.
[0] https://en.wikipedia.org/wiki/Global_Internet_usage
Isn’t there already a much older rule that predicts this?
Your product has to be a 10x improvement over the incumbent to be competitive.
In AI speak it would be the “extra-bitter” lesson I guess?
You need to add 10x resources to beat a product that’s already solved with mature tech.
is 3x worse like a 300% decrease?
Yeah, unclear what they mean. 1/3 the conversion rate, maybe?
> Walmart will embed its own chatbot, Sparky, inside ChatGPT. Users will log into Walmart, sync carts across platforms, and complete purchases within Walmart’s system.
The enshittification is upon us.
Hah, Clippy's cousin Sparky: every once in a while after ChatGPT answers a question it'll say "Looks like you still have stuff in your WalMart cart. Would you like me to complete that checkout for you? Also, WalMart-brand diapers is on offer this week, shall I add that to your cart?"
“You are an unfit mother.”
https://youtu.be/lajnHjRp9Z0
Original AI was sourced from university level text books, stack, wiki. Average iq around 140.
The latest AI is trained on the average citizens social media output. Iq 90.
That’s why AI seemed smart. The bar will not be raised again. We’re cooked.
source?
Because most people can't read. Wait for agents generating personalized websites/self checkout apps.
Not really many details...
Perhaps clickthrough is worse because there are fewer dark patterns involved and people are mostly just browsing and occasionally buying only what they need.
They didn't really seem to specify the "why" of it with any research. And weird that OAI wasn't supporting them to see wha the issue was.
What if they made instant checkout actually instant? You ask ChatGPT to setup a website and it instantly purchases web hosting and sets up the website there. You can't beat a 100% conversion rate by actually checking out instantly. If you didn't like that host you can ask it to find it alternative and ChatGPT would automatically attempt a refund and then purchase from someone else.
I'd become afraid to ask that bot anything, not knowing what it would automatically try to purchase for me. Paying for the bot itself is already a lot: $20/mo to $200/mo or more.
20$ for a personal assistant is not much, but no, I surely don't trust those assistants to access my money.
I pay $200/mo for one assistant, from one company. The number of other companies that want the same kind of money for other assistants is increasing. I wish I could pay $200/mo and get all the companies.
That's why it would be important for the partners to support refunding.
What are we talking about here? ChatGPT as an interface to buy groceries? Or ChatGPT as an interface to build a website. I fail to see how these would be related other than using a specific technology
> Because most people can't read. Wait for agents generating personalized websites/self checkout apps.
> You ask ChatGPT to setup a website and it instantly purchases web hosting and sets up the website
Multiple comments deflecting from the original shopping conversion failure to recommend ... building a whole new website (with hosting for some reason?). W/o bothering to look through commenter history, one has to assume there are a lot of chatbots on this site or else the people using this stuff have been lobotomized.
I'm sure it'll start happening too, and when that fails, the bots will, i don't know, invent a new macarena. We are definitely headed for an irredeemably stupid future.
I’m shocked, shocked that an SEO website would say ChatGPT bad.
Look, another thing AI is empirically bad at. Can we dispense with cramming AI into every product and service and just use it where it's useful? It's very wasteful and an utterly obnoxious experience for users.
The catalog normalization problem is real and severely underestimated. I ran into this building on top of product data feeds -- even with a single retailer, inconsistency in titles, attributes, and categorization is staggering. LLMs are great at reasoning but they inherit all that messy upstream data, so agentic shopping will keep stumbling until the data layer gets cleaner.